跳转到内容
Tessera Docs

Tessera Threat Model

此内容尚不支持你的语言。

Each threat is accompanied by a reference to its evidence: a path to code, the name of a config field, or a link to a test that proves the claimed property.

1. Introduction

Section titled “1. Introduction”

1.1 Target of evaluation (TOE)

Section titled “1.1 Target of evaluation (TOE)”

By tessera we mean:

  • the PAM module libpam_tessera.so (cdylib);
  • the tessera daemon (binary);
  • the core crate tessera_core and the protocol crate tessera_proto;
  • the shipped configuration files dist/config/*.example;
  • the systemd unit and the tmpfiles snippet;
  • the integration script dist/scripts/integrate-pam.sh.

The TOE does not include:

  • the Astra Linux SE kernel;
  • libpam, libssl3, libudev, libdbus, libsystemd;
  • gost-engine (a separate FSB-certified cryptographic module);
  • vendor PKCS#11 modules for tokens (separate FSB-certified cryptographic modules);
  • user hooks configured in [[hooks]];
  • Inkscape, fly-dm, gdm, and other PAM consumers.

2. Deployment assumptions

Section titled “2. Deployment assumptions”
#AssumptionWhy it matters
2.1The machine is physically protected (Astra’s mandatory integrity control, МКЦ, or equivalent).Without МКЦ, a root compromise → the module is useless.
2.2System integrity at install time is controlled (МКЦ + verified boot, if available).Replacing gost-engine.so or libpam_tessera.so bypasses the module.
2.3The CA infrastructure works correctly: keys in an HSM, issuance governed by policy.Compromise of the CA private key → catastrophic compromise of the whole perimeter.
2.4User PINs are not disclosed and not written on paper next to the computer.The PIN is the token’s only protection under physical access.
2.5Certificates are issued by the CA with the mandatory pam_cert_host_binding and pam_cert_allowed_roles extensions.Without the extensions, a certificate authorizes no role on any device — fail-closed.
2.6The administrator has a “backup tty” while editing the PAM stack.Protection against lockout on misconfiguration.
2.7A backup user with password authentication has not been removed.Lockout prevention if tessera fails.

3. Threats the module DOES protect against

Section titled “3. Threats the module DOES protect against”

Each threat follows the schema: description → STRIDE category → mitigation → evidence (code, config, test).

3.1 Password guessing

Section titled “3.1 Password guessing”
  • Description: an attacker tries to brute-force a local user’s password.
  • STRIDE: Spoofing.
  • Mitigation: tessera implements no password authentication. Any attempt to enter a password fails at the pam_conv stage.
  • Evidence: crates/pam_tessera/src/entry.rs contains no call to pam_authtok_get. Authentication proceeds via challenge-response with the private key (crates/tessera_core/src/challenge/).

3.2 Password leakage via shoulder-surfing / phishing

Section titled “3.2 Password leakage via shoulder-surfing / phishing”
  • Description: plain text password leakage.
  • STRIDE: Information Disclosure.
  • Mitigation: there is no password (see 3.1). The token PIN is not passed into the PAM stack or into logs.
  • Evidence: crates/tessera_core/src/secret.rs — the Secret<T: Zeroize> wrapper zeroes the PIN on Drop. The PIN is never passed as a format argument to a tracing macro (see also crates/tessera_core/src/pkcs12/).

3.3 Copying a certificate from USB media

Section titled “3.3 Copying a certificate from USB media”
  • Description: an attacker copies a .p12 from someone else’s USB and tries to use it on their own machine.
  • STRIDE: Spoofing.
  • Mitigation (Mode A): the .p12 is encrypted with a passphrase; without the passphrase decryption is impossible. Mode A is considered a “medium” protection mode — production uses Mode B.
  • Mitigation (Mode B): the key is non-extractable. The test checks the CKA_EXTRACTABLE = false attributes.
  • Evidence: test crates/tessera_core/tests/pkcs11_hardware_negative.rs

3.4 Using someone else’s token without knowing the PIN

Section titled “3.4 Using someone else’s token without knowing the PIN”
  • Description: the attacker obtained the token (stole it / picked it up from a chair) but does not know the PIN.
  • STRIDE: Spoofing.
  • Mitigation: PIN prompt via the PAM conversation; after N failed attempts (pkcs11_max_pin_attempts, default 3) the module denies access. After N attempts on the token itself, it is locked at the level of the hardware counter.
  • Evidence: test crates/tessera_core/tests/pin_loop.rs checks the attempt limit.

3.5 Using a valid token on someone else’s machine

Section titled “3.5 Using a valid token on someone else’s machine”
  • Description: the attacker legitimately owns the token (or stole it) but tries to use it on a machine where this token is not permitted.
  • STRIDE: Spoofing + Elevation of Privilege.
  • Mitigation: verification of the X.509 pam_cert_host_binding extension — its entries (* / sha256:<HEX> / raw machine_id) are compared against the host_id_hash = sha256(host_id) of the requesting machine. If no entry matches — PAM_AUTH_ERR (HostNotAllowed). The extension itself is protected by the CA’s certificate signature — it cannot be changed without compromising the CA.
  • Evidence:

3.6 Using a session after the user leaves

Section titled “3.6 Using a session after the user leaves”

3.6.1 Astra ЗПС (DIGSIG) and binary signing

Section titled “3.6.1 Astra ЗПС (DIGSIG) and binary signing”
  • Description: an attacker with file-write privileges replaces pam_tessera.so or tessera with a forged binary.
  • STRIDE: Tampering.
  • Mitigation:
    • On Astra Linux SE the production mode is astra-digsig-control in enforce. ELF files from the tessera package must be signed by the build CI of the Astra partner (bsign with a GPG key from the trusted keyring /etc/digsig/keys/); a substituted binary without a corresponding signature is rejected by the kernel at execve(2) / mmap(2).
    • If ЗПС — the closed software environment (Astra’s signed-executables enforcement) — is switched to logging-only, protection drops to dpkg --verify and 0755 root:root permissions on the binary — this mode is acceptable only on dev machines. A production deploy without signing is forbidden by operational policy.
  • Evidence: see install.md §1.5 “Preflight: USBGuard and Astra ЗПС (DIGSIG)” — it describes both the enforcement mode and the diagnostic commands.

3.7 Private-key leakage from process memory

Section titled “3.7 Private-key leakage from process memory”
  • Description: an attacker reads process memory (via ptrace, /proc/pid/mem, a minidump) and tries to extract the PIN or the private key.
  • STRIDE: Information Disclosure.
  • Mitigation:
    • The PIN is stored in Secret<T: Zeroize> — zeroed on Drop.
    • In Mode B the private key never leaves the token (PKCS#11 non-extractable).
    • In Mode A the decryption passphrase is used once and zeroed after flow::authenticate.
    • The systemd unit sets NoNewPrivileges=yes, ProtectKernelTunables=yes, RestrictNamespaces=yes — making ptrace from outside harder.
  • Evidence:

3.8 Certificate without extensions or with a forged extension

Section titled “3.8 Certificate without extensions or with a forged extension”
  • Description: an attacker tries to use a certificate that has no pam_cert_host_binding / pam_cert_allowed_roles extensions at all, or attempts to embed “forged” entries bypassing the CA.
  • STRIDE: Tampering + Spoofing.
  • Mitigation:
    • Mandatory-extension policy: the absence of either extension in the leaf certificate is an unconditional denial: no host extension → HostExtensionMissing → PAM_AUTH_ERR; no allowed_roles → the credential grants no role at all → PAM_PERM_DENIED. There are no “soft” fallbacks, and no device-side admission exists: the scope is assigned by the issuer.
    • Protection by the CA signature: the extension content is covered by the certificate signature. Changing an entry without the CA private key is impossible; forging the whole certificate is the task of compromising the CA (see 4.8).
    • Chain verification: issuing a certificate by a rogue “trusted” CA is caught by [trust].anchors + optionally [trust.pinning].
    • Corrupted DER encoding (garbage in extnValue) → a fail-closed denial: for the host extension *ExtensionMalformed → PAM_AUTH_ERR; for allowed_roles the list is treated as empty (not ignored) → the requested role is not covered → denial.
  • Evidence:
    • implementation — tessera_core::x509::{host_binding_ext, allowed_roles_ext} + verify_cert_scope;
    • semantics table — docs/cert-issuance.md.

3.9 Substituting config.toml

Section titled “3.9 Substituting config.toml”
  • Description: an attacker with file-write privileges replaces config.toml with a configuration that has a weakened [trust] or disabled revocation.
  • STRIDE: Tampering.
  • Mitigation:
    • config.toml is not signed but is protected by 0640 root:root permissions (see debian/postinst);
    • dpkg --verify tessera detects changes to the shipped files (but not user edits to config.toml after installation);
    • changing the config requires root access — which is already outside the PAM-level threat model.
  • Evidence: debian/postinst + dist/tmpfiles/tessera.conf.

3.10 MITM in IPC

Section titled “3.10 MITM in IPC”
  • Description: an attacker tries to connect to /run/tessera/monitord.sock and substitute responses.
  • STRIDE: Tampering + Spoofing.
  • Mitigation:
    • the socket in /run/tessera/ has 0660 root:tessera permissions (see systemd RuntimeDirectoryMode=0750);
    • monitord checks the peer via SO_PEERCRED — uid != 0 → Error { code: 1003 (UNAUTHORIZED) } + disconnect.
  • Evidence:

3.11 Replay attacks on challenge-response

Section titled “3.11 Replay attacks on challenge-response”

3.12 Argv injection in hooks

Section titled “3.12 Argv injection in hooks”

3.13 Weak-signature-algorithm attack

Section titled “3.13 Weak-signature-algorithm attack”

4. Threats the module does NOT protect against

Section titled “4. Threats the module does NOT protect against”
#ThreatRecommended compensating control
4.1Rootkit / kernel compromiseAstra МКЦ, IMA, EDR.
4.2Physical access to an unlocked session before grace fires.Reduce usb_removed_grace_seconds; admin policy.
4.3Extracting the key from the token given a compromised PIN + physical access to the token + special equipment.The token’s hardware measures (in-chip anti-tamper).
4.4Mode A with a .p12 without a password or with a weak password.Do not use Mode A in production. Use Mode B.
4.5Vulnerabilities in gost-engine.Timely Astra updates; the FSB-certified module is responsible for patches.
4.6Side-channel attacks on the token (electromagnetic, power, timing).Hardware anti-tamper measures.
4.7Social engineering (handing over the token and disclosing the PIN).Training, policy.
4.8CA compromise (CA private key leaked).HSM for the CA, separation of roles; [trust.pinning] limits the blast radius to the pinned-roots level.
4.9Vulnerabilities in libpam, libssl3, the kernel.apt updates, CVE monitoring.

4.10. Host identity: multi-source matching is NOT performed

Section titled “4.10. Host identity: multi-source matching is NOT performed”

[host_identity].sources is defined as an ordered fallback, not as “accept if it matched any of the sources”. resolve() takes the first successful source and computes only its host_id_hash; the cert must encode exactly this value (via pam_cert_host_binding).

This is intentional. Multi-source “matched at least one” matching is equivalent to “weakest source wins”: an attacker with root privileges substitutes the most writable source (for example /sys/class/dmi/id/board_serial under qemu, or custom_command via a shim), and host binding is bypassed. Therefore matching is checked ONLY against the resolved source per the fallback policy.

Effect for administrators: after changing [host_identity].sources, the cert must be re-issued (new source → new host_id_hash). Drift between the issuance script and the deployed configuration is caught via journalctl -t tessera | grep 'host_id resolved' (see install.md).

5. Attack surface

Section titled “5. Attack surface”
#SurfaceProtection
5.1PAM stack (/etc/pam.d/*)Standard PAM security; integrated via @include tessera.
5.2libssl3 / libcryptoSystem updates via apt.
5.3PKCS#11 module (Rutoken / JaCarta)FSB-certified cryptographic module; closed-source; trusted while its certification is valid.
5.4udev eventsNot authenticated, but we are inside the kernel namespace and trust udev.
5.5IPC socket /run/tessera/monitord.sockSO_PEERCRED uid=0 + 0660 permissions. If root is already compromised, the module is already useless.
5.6Hooks in [[hooks]]Placeholder whitelist, fork+execve, timeouts. The hook itself is the administrator’s responsibility.
5.7Config files /etc/tessera/config.toml and trust anchors0640 root:root permissions. Manual management.
5.8Backend plugins /usr/lib/tessera/plugins/*.soEd25519 signature against a build-embedded key list before dlopen (non-disableable in shipped builds); RTLD_NOW+RTLD_LOCAL; strict ABI validation; any failure → StubBackend + audit. Details in §5.6.

5.1 Process privilege model

Section titled “5.1 Process privilege model”
ProcessContext / UIDHardeningKnown residual risk
pam_tessera.soUID of the PAM caller (sudo/login/fly-dm — usually root at the auth stage); an architectural requirement of PAM.#![forbid(unsafe_code)] on tessera_proto; panic_guard at every C boundary → PAM_AUTHINFO_UNAVAIL; Secret<T: Zeroize> for the PIN.Loading into the address space of a rooted process — compromise of the host compromises the module too (outside the TOE, see 4.1).
tesseraUser=tessera / Group=tessera — a dedicated system account without a shell, created by debian/postinst.ProtectSystem=strict + ReadWritePaths=…, ProtectHome=yes, PrivateTmp=yes, NoNewPrivileges=yes, ProtectKernelTunables/Modules/ControlGroups=yes, RestrictNamespaces=yes, RestrictRealtime=yes, LockPersonality=yes, CapabilityBoundingSet=CAP_DAC_READ_SEARCH, AmbientCapabilities=CAP_DAC_READ_SEARCH. Privileged D-Bus calls to logind are gated by a polkit rule.MemoryDenyWriteExecute=no (left off due to the W^X relaxation in OpenSSL/gost-engine); full W^X sandboxing is a task for after the benchmarking stage (see the systemd unit and the backlog for 0.1.2).

pam_tessera.so runs in the context of the PAM caller — this is an architectural constraint of the PAM stack, not an implementation choice; the cdylib’s privileges cannot be lowered without redesigning the PAM protocol. As of 0.1.1, tessera is already split into a separate system account — root privileges for D-Bus actions on logind are granted narrowly via a polkit rule shipped with the package.

5.2 Lockout-resilience model

Section titled “5.2 Lockout-resilience model”

The PAM stack into which tessera is integrated turns the USB token into a hard second factor (or the only factor, see cert-only). This is a deliberate security choice; the price of that choice is that resilience to token loss falls on operations, not on the module itself:

ModeToken lossUSBGuard blocks the tokenAstra ЗПС in enforce without a signed binary
2faYou can log in with a password.Same — the password works.The PAM module will not load → fallback to the password (auth required will break the login).
optionalYou can log in with a password.Same.Same.
cert-onlyLockout. Even local root cannot get in.Lockout.Lockout — auth [success=done default=die].

Compensating controls for cert-only (mandatory before deploy):

  • a backup access channel without tessera (see install.md §9) — a separate sshd stack with UsePAM=no or a sudoers rule for an emergency account;
  • a spare token with the same pam_cert_allowed_roles for every privileged user;
  • a documented rescue-recovery procedure (see install.md §11 “Lockout after a failed PAM edit”).

5.6 Enforcement backend plugins

Section titled “5.6 Enforcement backend plugins”

Platform enforcement backends (ParsecBackend for Astra МКЦ; SELinux later) are delivered not as compile-time features but as signed runtime plugins: a single open binary loads the backend from a separate .so. This removes the build combinatorics but introduces a new surface — execution of third-party code inside the address space of a rooted process (pam_tessera during the auth phase, see 5.1). Below is what closes that surface.

Code: crates/tessera_core/src/plugin/ (verify.rs — signature, loader.rs — loading and the C-ABI ↔ MacBackend bridge, header.rs — envelope/vtable, audit.rs — events).

5.6.1 Trust: detached signature before dlopen

Section titled “5.6.1 Trust: detached signature before dlopen”
  • Algorithm (as implemented): Ed25519 detached signature over the exact bytes of the .so. Signature file <plugin>.so.sig, format ed25519:<128 hex>. The algorithm prefix is mandatory — an old signature cannot be reinterpreted once a future ABI adds GOST.
  • Trust anchor: the public-key list is embedded at build time (TESSERA_PLUGIN_PUBKEYS — 32-byte raw Ed25519 keys in hex). An empty list → NoKeys: every plugin is rejected (fail-closed).
  • Order (loader.rs): is_file → signature verification → read and SHA-256 (for audit) → dlopen. The signature is checked before any plugin code; a foreign key or an altered image yields Invalid. Caveat: signature, SHA-256 and dlopen are three separate reads of the file by path, so “an altered image does not load” strictly holds only without a concurrent swap between those reads (TOCTOU, see 5.6.5).
  • Verification is non-disableable in shipped builds. It is skipped only under cfg(debug_assertions); the .deb is built with the release profile (no debug-assertions), and a debug build does not ship. An atypical release with debug-assertions enabled would bypass the check — the package smoke test rejecting an unsigned plugin guards against that.

5.6.2 Load isolation

Section titled “5.6.2 Load isolation”
  • dlopen with RTLD_NOW | RTLD_LOCAL: unresolved symbols are rejected immediately (no lazy binding), and plugin symbols do not leak into the process global namespace.
  • ABI-contract validation of the envelope: abi_version == 1, kind == enforcement, non-null name/version/vtable, and the header name matches the name explicitly selected by config. Field mismatches and null pointers → rejection. However, the header is parsed through pointers (&*header_ptr, CStr::from_ptr, vtable cast), so a signed but broken/version-diverged header carrying a non-null yet invalid pointer is UB during parsing, before the panic boundary (residual risk, see 5.6.5). The bar here is trust in the vendor signature.
  • Panic boundary (plugin contract): every callback must catch panics and return PLUGIN_PANIC before crossing the C ABI (unwind cannot safely cross independently linked dylib runtimes); the host does not force this, it relies on the contract. PLUGIN_PANIC → audit and degradation, not a process crash. Exception — teardown: it has no status code, and its panic is not logged.
  • An accepted plugin stays loaded for the process lifetime (ManuallyDrop) — dlclose is not called for the active backend, no dangling pointers. On failure paths after dlopen (entry/header/abi/kind/name/init) the library is unloaded by a normal dlclose; on an init failure teardown is not called.

5.6.3 Selection and degradation

Section titled “5.6.3 Selection and degradation”
  • The active backend is named explicitly by config ([mac] backend = "parsec"); there is no auto-activation. Without a name — StubBackend.
  • Any failure (missing file, signature, dlopen, envelope, ABI, init) → plugin_rejected → degradation to StubBackend. Fail-closed for roles requiring МКЦ is enforced not by the loader but by role policy (change role-format): StubBackend issues no МКЦ label, and a role demanding integrity is rejected.

5.6.4 Audit (target = plugin.audit)

Section titled “5.6.4 Audit (target = plugin.audit)”
EventWhenFields
plugin_loadedplugin acceptedname, plugin_version, kind, sha256
plugin_rejectedany load failurepath, reason (missing/signature/dlopen/header/abi/kind/init)
plugin_inactive_file.so present in the directory but not selectedpath
plugin_paniccallback returned PLUGIN_PANICname, entry_point

plugin_loaded is emitted right after init — before the final probe — so it means “initialized candidate”, not “the backend became active” (the runtime mode may still degrade to StubBackend). The SHA-256 in the event is taken from a separate file read, not from the in-memory image (see TOCTOU in 5.6.5) — a forensic pointer to the candidate, not a cryptographic binding to the active image.

5.6.5 Residual risks

Section titled “5.6.5 Residual risks”
RiskAssessment
.so swap by local rootDirectory is root:root; write = root already compromised (outside the TOE, 4.1). The signature is defense-in-depth: even root cannot substitute its own backend without the vendor private key, short of rebuilding the binary (itself under dpkg/ЗПС signing, see 3.6.1).
Build without embedded keysA release mistake → NoKeys → enforcement disappears, but fail-closed + plugin_rejected. Closed by the package smoke test.
Debug build without verificationOutside the shipping model (cfg(debug_assertions), not in the .deb).
Ed25519 signature, not GOSTThe plugin signature is vendor delivery infrastructure, not a certificate chain; Ed25519 is acceptable. The mandatory format prefix keeps a GOST migration unambiguous.
Malicious/buggy plugin after loadRuns with the process privileges. The boundary is the signature (trusted vendor) and the panic guard (cooperative, plugin-side); the plugin is thread-safe by contract.
Broken/version-diverged signed headerA non-null but invalid pointer in the header → UB during parsing (&*header_ptr/CStr::from_ptr/vtable cast) before the panic boundary. The bar is trust in the vendor signature; envelope typing only catches field mismatches and null.
TOCTOU: signature↔SHA-256↔dlopenThree separate reads of the file by path; a concurrent writer could substitute a different image between verification and dlopen. Requires writing to a root:root directory = root already compromised (outside the TOE, 4.1); binding to a single immutable read (memfd) is a possible future measure.

6. Adversary model

Section titled “6. Adversary model”
LevelDescriptionModule’s expectation
A1External adversary without physical access, without a token, without a PIN.Zero successes.
A2External adversary who obtained a token but does not know the PIN.Zero (PIN protection + attempt limit).
A3Internal adversary: a legitimate user tries to use the token on a forbidden machine or to log into another role account.Zero (host_binding + allowed_roles in the extensions, protected by the CA signature).
A4Internal adversary: an administrator with root access.Not modeled — the admin is trusted by construction.

7. Attack tree for threat 3.5 “using a valid token on someone else’s machine”

Section titled “7. Attack tree for threat 3.5 “using a valid token on someone else’s machine””
graph TD
    R[Get a shell as alice<br/>on terminal-XYZ]
    R --> A1[Steal alice's token<br/>+ learn the PIN]
    R --> A2[Forge a certificate<br/>with CN=Alice]
    R --> A3[Obtain from the CA a certificate<br/>with an extension permitting terminal-XYZ]
    R --> A4[Spoof host_id<br/>to impersonate a permitted machine]


    A1 --> A1a[Social<br/>engineering]
    A1 --> A1b[Theft + observation]
    A1a -.->|compensating| C1[User training]
    A1b -.->|compensating| C2["'token with you' policy"]


    A2 --> A2a[Compromise the CA]
    A2 --> A2b[Find a hash collision]
    A2a -.->|mitigation| M1[CA in an HSM<br/>+ trust.pinning]
    A2b -.->|mitigation| M2[Whitelist of strong algorithms<br/>3.13]


    A3 --> A3a[Certificate without extensions]
    A3 --> A3b[Compromise the CA]
    A3a -.->|mitigation| M3[Mandatory-extension policy<br/>3.8]
    A3b -.->|mitigation| M1


    A4 --> A4a[machine_id<br/>matched by chance]
    A4 --> A4b[Spoof /etc/machine-id]
    A4 --> A4c[Spoof DMI board serial]
    A4a -.->|mitigation| M4[host_id based on several sources]
    A4b -.->|mitigation| C3[File integrity under МКЦ]
    A4c -.->|mitigation| C4[Impossible without opening the case]

8. List of tests proving the claimed protections

Section titled “8. List of tests proving the claimed protections”
ThreatTest(s)File
3.3non-extractable checkcrates/tessera_core/tests/pkcs11_hardware_negative.rs
3.3PKCS#12 wrong passwordcrates/tessera_core/tests/pkcs12.rs
3.4PIN attempt limitcrates/tessera_core/tests/pin_loop.rs
3.5host_binding mismatch (sha256 entry did not match)crates/tessera_core/tests/verify_cert_scope.rs
3.5end-to-end auth with host/user binding extensionscrates/pam_tessera/tests/auth_e2e_p12.rs
3.6USB removal → grace → lockcrates/tessera_cli/tests/udev_simulation.rs
3.6suspend/resume ignores transient REMOVEcrates/tessera_cli/tests/suspend_grace.rs
3.7Secret zeroization on Dropunit tests in crates/tessera_core/src/secret.rs
3.8certificate without extensions → denialcrates/tessera_core/tests/verify_cert_scope.rs (negative cases)
3.10uid≠0 peer is rejectedcrates/tessera_cli/tests/peercred.rs + ipc_auth.rs
3.11challenge is not repeatedcrates/tessera_core/tests/challenge_dispatch.rs
3.12argv injection is impossiblecrates/tessera_core/tests/hook_security_integration.rs
3.13weak signature → DisallowedSignatureAlgorithmcrates/tessera_core/tests/chain_verify.rs
3.13GOST chain verify (real engine)crates/tessera_core/tests/gost_chain_verify_real.rs
Reproducibility / supply-chainreproducible build (double build)scripts/verify-reproducible-build.sh

9. МКЦ (Astra strict mode, 0.3.0+)

Section titled “9. МКЦ (Astra strict mode, 0.3.0+)”

9.1 Threats

Section titled “9.1 Threats”
  • 9.1.1 Privilege-escalation via MAC label. The certificate declares an excessively high MAX_INTEGRITY; without runtime control the user raises the session level above the host’s ceiling.
  • 9.1.2 Bypass through missing extension. A certificate issued before МКЦ was deployed does not contain MAX_INTEGRITY — without cert_integrity = "required" the session opens without a label and gets “transparent” access.
  • 9.1.3 DER-tampering. An attacker controlling the CA puts broken/non-standard DER into the extension, counting on a parser failure and “accept-by-default” fallback behavior.
  • 9.1.4 sessions.json TOCTOU. The state file is overwritten atomically. Previously the irelax label on a new inode was restored by a separate syscall → a race window in which the daemon with MAC=0 cannot read the just-written file. Fixed in 0.3.0: the file lives on the tmpfs /run/tessera/ (RuntimeDirectory=), the parent directory gets iinh, the label is applied to the fd before the name is published via pdp_set_fd (see §9.2.4); a reboot clears the state entirely.
  • 9.1.5 host_id rebind. By spoofing host_id, the attacker rebinds the certificate to a different host.

9.1.6 irelax + UID 0 = forge (in-scope, out-of-mitigation)

Section titled “9.1.6 irelax + UID 0 = forge (in-scope, out-of-mitigation)”

monitord and pam_tessera.so apply irelax to their own files (/run/tessera/monitord.sock, /run/tessera/sessions.json) via the PARSEC_CAP_CHMAC privilege of the tessera user. This is necessary: an engineer at integrity level 1 (integrity mask = 1) must be able to write receipts into the lvl=0 daemon through a socket that is itself labeled irelax.

A hostile UID-0 process (root-equivalent) with the same PARSEC_CAP_CHMAC capability can attach identical irelax labels to its own files and forge entries in sessions.json, or connect to the socket from an arbitrary integrity level. Not blocked in code: the defense is the UID boundary (root-equivalence axiomatically wins under МКЦ), not integrity.

Mitigations beyond the scope of this threat:

  • DAC 0600 root:root on sessions.json + parent dir 0750 tessera:tessera restrict write access for non-CHMAC processes.
  • digsig verification on the tessera(-monitord) binaries detects runtime tampering.
  • An audit log entry on every mac_apply_failed / mac_caps_missing exposes unexpected backend behavior.

The trust boundary of this design is UID 0 vs non-root, not the integrity level. The threat is acknowledged as in-scope and explicitly out-of-mitigation for the PAM layer: protection against UID-0 forgery is the OS’s job (digsig, ЗПС, a hardware root of trust), not the authentication module’s.

9.2 Protections

Section titled “9.2 Protections”
  • 9.2.1 The effective label is always intersected with the user’s ceiling from the backend (the user integrity ceiling, МНКЦ, get_user_mnkc) — the ceiling is set by the system, not the certificate. See the free function compute_effective_label (mac/orchestrator.rs).
  • 9.2.2 cert_integrity = "required" rejects certificates without the extension; the stub backend refuses to start with required.
  • 9.2.3 The IntegrityLabel::from_der parser is strict: length checks, no trailing bytes, BIT STRING unused-bits ≤ 7. Broken DER → denial + the audit event cert_max_integrity_parse_failed.
  • 9.2.4 Writing sessions.json goes through openat(O_TMPFILE) → fchmod → pdp_set_fd(label) → linkat/rename atomically; the label is applied to the fd before the name is published. The kernel does not accept irelax through the fd-based API (EINVAL) — the relax semantics for sessions.json are provided by iinh inheritance from the parent dir /run/tessera/ (tmpfs).
  • 9.2.5 postinst applies chattr +i to host_id after the first write; the file itself lives in the dir /var/lib/tessera/ (0750 root:tessera). Daemon-writable persistent state is isolated under /var/lib/tessera/daemon/ (0750 tessera:tessera), so a daemon compromise cannot replace trusted role/tag paths through their parent.

9.3 Open risks

Section titled “9.3 Open risks”
  • The libparsec parsec_capget symbol soname is not fixed publicly; build.rs does not link libparsec-base3 by default. If, on a particular Astra release, the build produces an “undefined symbol”, you need to add libparsec-base3 to debian/control and cargo:rustc-link-lib=parsec-base to build.rs.
  • libpdp.so.3 is proprietary, without public fuzzing coverage. Protection: LD_LIBRARY_PATH is fixed, the ABI is wrapped in a return-pointer signature (see 5337fea), all calls run under panic=abort.

9.4 Tests

Section titled “9.4 Tests”
ThreatTestFile
9.1.1intersect(cert, caps) capping levelcrates/tessera_core/tests/mac_orchestrator.rs
9.1.2cert_integrity=required rejects no-ext leafcrates/pam_tessera/tests/mac_open_session.rs::open_session_fails_when_required_but_cert_lacks_ext
9.1.3malformed DER → parse-failed eventcrates/tessera_core/tests/cert_extensions_parse.rs
9.1.4fd-based irelax label on atomic writecrates/tessera_core/tests/mount_guard_tmpfs.rs
9.1.5host_id immutability after installE2E manual: vagrant/scripts/test-mac.sh T12

10. Threat registry (systematic pass 2026-06)

Section titled “10. Threat registry (systematic pass 2026-06)”

The result of a full bootstrap pass over the code @ 14b828e (a research swarm over docs/surfaces/assets/infrastructure/git-history, then clustering and STRIDE gap-fill) refined with the owner. The format is compatible with the THREAT_MODEL.md schema (downstream tools parse the table by regex). A threat is a class that survives the patch of a specific bug; the bugs found are listed in evidence and only raise likelihood.

A static triage pass (2026-06-06, 25 raw findings → 5 confirmed) verified instances of this registry. An important axis: impact/likelihood in the registry are an estimate of the class of threat in the worst-case scenario; a specific instance, given a stack of preconditions, may be instance-severity LOW while the class-impact is critical. Both numbers are needed: the class drives investment priority in mitigations, the instance drives patch urgency. The pass also uncovered two classes absent from the bootstrap: false revocation via issuer scope (added to T1/T8) and path confusion on mounted media (the new T12).

This section complements §3–§6: §3 describes protection mechanisms per threat; here is a single ranked registry (impact × likelihood, descending).

10.1 Threats

Section titled “10.1 Threats”
idthreatactorsurfaceassetimpactlikelihoodstatuscontrolsevidence
T1Login with a revoked credential: residual revocation degradation (an expired CRL is skipped when crl_strict=false — “revocation is not forever”; a CRL without nextUpdate is bounded only by the opt-in crl_max_age_hours)insiderCRL processingAuthentication decision, revocation statecriticalpossiblepartially_mitigatedshort credential TTL limits the window; the CRL signature is verified mandatorily (fail-closed, crl/store.rs); OCSP is implemented (ocsp/crl_then_ocsp, nonce, issuer-signer check); an empty CRL store with mode=crl is rejected at verifier construction (2026-07); crl_strict=true is opt-in; issuer-DN binding in check_revocation (RFC 5280 §6.3.3) — a CRL applies only to certificates from its own issuer14b828e, openspec/revocation, F-001/F-002 (closed)
T2Bypassing the authorization policy via fail-open defaults and silent fallback pathsinsiderChain verification and challenge-response; X.509/PKCS#12 parsing from media; config.tomlAuthentication decision, fail-closed invariantcriticalpossiblepartially_mitigatedmandatory-extension policy (host_binding/allowed_roles); the silent admission fallback path was eliminated in 2026-07: the pam_cert_user_binding extension and the [[user_mapping]] section were removed, admission is decided solely by allowed_roles of the verified credential, and a config carrying the removed section is rejected at validation; strict DER parsing of МКЦ labels; since 2026-06 an empty/omitted sig-whitelist is replaced by a safe default (SHA-256/384/512 RSA + ECDSA, without SHA-1/GOST) during config validation — the accept-all default is eliminated; since 2026-06 an extractable PKCS#11 key is rejected by default (ExtractableKeyRejected), a WARN remains only under the explicit opt-in pkcs11_allow_extractable_keys = true; since 2026-08 the paired fail-open is closed as well: provider silence is no longer read as CKA_EXTRACTABLE = FALSE — a token that did not report the attribute is rejected with its own ExtractableAttributeUnavailable error and its own opt-in pkcs11_allow_unreported_extractable, because cryptoki silently drops from the read result any attribute the token declined to return; since 2026-07 omitting the [trust.revocation] section or the mode key is a config validation error (the silent “revocation not checked” default is eliminated; none must be chosen explicitly)pre_validate.rs:28, flow.rs:662, key_lookup.rs, config/validated.rs, F-002 (closed)
T3RCE/memory corruption in the root login process while parsing malicious media: DER/PKCS#12 in OpenSSL before verification, and the filesystem image in the kernel at mount(2)local_userX.509/PKCS#12 parsing from media; USB mountHost process integritycriticalpossiblepartially_mitigatedRust wrapping; panic guard (does not save from UB in C); mount with nosuid,nodev,noexec; the CVE history of ASN.1 parsers / filesystem drivers as precedent
T4Evil-maid outside Astra: substitution of config.toml, native .so files (PKCS#11/gost-engine), host_id — without МКЦ labels, DIGSIG, and the immutable bit, Debian/Ubuntu are protected only by DAClocal_userDynamic loading of native code; config.toml; Host identity resolution; Package install/removalHost process integrity, configuration, host identitycriticalpossiblepartially_mitigatedon Astra: МКЦ ilevel=63, chattr +i, DIGSIG/ЗПС; outside Astra: 0640/0750 root:tessera
T5Build/supply-chain compromise: a backdoor in the login-stack .so across the whole fleet via an unpinned builder-image base, tools without checksums, unsigned .deb, a crates.io dependencysupply_chainCI / build supply chain; Cargo dependenciesIntegrity of release artifacts, Host process integritycriticalpossiblepartially_mitigatedreproducible build (.buildinfo), rust-cache pinned by hash, cargo-deny, draft releases; Astra DIGSIG will reject an unsigned .so; GPG signing of .deb and an apt repository are planned
T6Memory-safety violation at the PAM FFI boundary: manual ownership of AuthContext (Box::into_raw), conv pointers, unsafe blocks in the cdyliblocal_userPAM ABI (pam_sm_*); PAM data (AuthContext between phases)Host process integritycriticalrarepartially_mitigatedRust; panic_guard → PAM_AUTHINFO_UNAVAIL; forbid(unsafe) in proto; no-alloc between fork/execve
T7The session survives media removal: the monitoring path is not fail-closed (a SessionOpen error is not fatal under strict)local_usermonitord IPC socket; udev events; sessions.json persistenceActive session, media-removal controlhighlikelypartially_mitigatedudev REMOVE → grace → lock/logout/hook/shutdown; race-check on SessionOpen; fix of the XDG_SESSION_ID path in 0.3.13; since 2026-06 Lock/Logout without a logind id is fail-closed: an error log + host reboot (the session is destroyed, the machine returns to the login screen) instead of silently dropping the actionflow.rs:742, actions.rs:51, F-006 (closed)
T8Device unavailability (DoS/lockout): cert-only without a rescue channel, deliberate blocking of someone else’s token (3 PIN attempts), USB timeout, monitord strict, false expiry on clock drift, false revocation of a valid credential on a cross-CA serial-number collision (issuer-DN not checked, F-001)local_userPAM ABI (pam_sm_*); monitord IPC socket; udev events; CRL processingAuthentication decision (availability)highlikelypartially_mitigated2fa/optional modes with password fallback; a rescue channel for cert-only is documented but not enforced; grace windows; clock_skew tolerance in acct_mgmt (lib.rs acct_mgmt_core)F-001
T9Incorrect PAM stack after install/upgrade/removal: bypass (leftovers of optional mode, uncleaned @include) or full lockoutlocal_adminPackage install/removal (postinst/postrm, integrate-pam.sh, finish-bootstrap.sh); PAM ABIThe system PAM stack, Authentication decisionhighpossibleunmitigatedbackup bak.; interactive y/N in finish-bootstrap; there is no syntactic PAM validation after editing
T10Privilege escalation via the hook mechanism: PAM-derived variables in the child process’s environmentlocal_userHooks fork+execveHost process integrity, active sessionhighrarepartially_mitigatedplaceholders as separate argv (no shell), setuid/setgid in the child, timeouts; since 2026-06 a pre-exec check of the permissions of the hook file and all parent directories (like sudo/ssh): S_IWOTH — always denied, S_IWGRP — denied except for the root/egid group, an unreadable stat — denied (fail-closed); the residual TOCTOU window stat→execve is equivalent to sudo/sshF-004 (closed), validator.rs:71, child_setup.rs:335
T11Reconnaissance via state disclosure: the session registry (who is logged in), diagnostics on the login screen, the audit loglocal_usersessions.json persistence; PAM ABISession registry, audit eventslowpossiblepartially_mitigated0660/0750, tmpfs; the “issued for another device” diagnostic is a deliberate UX trade-off
T12Path confusion on mounted USB media: a symlink in ext4 (or another symlink-capable FS) on the USB → reading host files as root; MS_NOSYMFOLLOW is not set in the mount flagsphysical_userUSB mount; tessera_core/discovery.rsConfidentiality of host files (partial)mediumrarepartially_mitigatedsince 2026-06 discover_credentials does canonicalize + boundary check (starts_with the canonical mountpoint) for p12 and chain.pem; escaping the mountpoint → denial (EscapesMount, fail-closed); this closes both mid-path symlinks and ..; MS_NOSYMFOLLOW is still not set in the mount flags (defense-in-depth, open)F-016 (closed), discovery.rs:92, usb.rs:85

10.2 Deprioritized

Section titled “10.2 Deprioritized”
threatreason
Spoofing/Tampering of network channelsDevices are offline, there are no network login interfaces
Root on the device, full offline disk accessOut of the model by construction — the environment integrity layer (see §2, §4.1)
Repudiation of multi-user actionsPartially covered by structured audit; correlation with the server-side session is the server side’s job, not the PAM module’s
Forging udev eventsRequires root or CAP_NET_ADMIN on netlink — equivalent to a root adversary
Side-channel / hardware attacks on the tokenA property of the certified token (FSB-certified crypto), outside the TOE (see §4.3, §4.6)
Timing attacks on challenge-responseThe signature is performed by OpenSSL/the token; no secret-dependent comparisons were found in the tessera code
DoS by physical damage to the deviceOut of the model (physical protection of the environment)
Section titled “10.3 Recommended mitigations (class-level)”
mitigationthreat_idscloses_classeffortstatus
New CRL semantics: revocation is permanent for serial numbers from an expired CRL; expiry affects only trust in the completeness of the list + an audit eventT1yesMpartially closed 2026-06/07: verify_signature is mandatory, OCSP is implemented, an empty store with mode=crl is fail-closed; ONLY “revocation is permanent for an expired CRL” remains OPEN (a stale CRL is skipped when crl_strict=false)
Fail-closed config invariant: an empty sig-whitelist = a validation error; a missing revocation mode = a validation error; a malformed binding extension = denial (not fallback); CKA_EXTRACTABLE=true = block under hardware policyT2yesSsig-whitelist implemented 2026-06 (variant: a safe default instead of a validation error — does not break existing configs); revocation mode required since 2026-07 (omitting the section/key = a validation error, the silent none default eliminated); EXTRACTABLE implemented 2026-06; binding fallback — open
Align the code with the docs in monitoring: strict → a SessionOpen error is fatal; a session without a logind id degrades to a tty target instead of dropping the actionT7yesSthe logind-id part implemented 2026-06 (variant: fail-closed reboot instead of tty degradation); strict→fatal — open
Sign .deb (GPG) + pin the builder base image by digest + checksums for curl-downloaded toolsT5yesSopen
Post-validation of the PAM config after each edit (syntax check, smoke test) + atomic rollback in integrate-pam.shT9yesMopen
Debian integrity profile: dpkg-statoverride on critical paths, AIDE/IMA recommendation, checking hook-file permissions at daemon startupT4, T10partialMhook permissions implemented 2026-06 (variant: pre-exec check in fork_exec, stricter than at daemon startup); the rest of the profile — open
Isolate media parsing: a privilege-separated helper for PKCS#12/DER before verification; size limitsT3partialLopen (p12/chain limits exist)
Check the rescue channel in finish-bootstrap.sh for cert-only (clock_skew tolerance in acct_mgmt — done)T8partialSopen
MS_NOSYMFOLLOW in mount(2) + canonicalize/boundary check in discover_credentials before fs::readT12yesScanonicalize/boundary check implemented 2026-06; MS_NOSYMFOLLOW — open (defense-in-depth)
CRL issuer-DN binding in check_revocation: compare the cert’s issuer_dn_der with the CRL issuer before matching by serial number (RFC 5280 §6.3.3)T1, T8yesSimplemented 2026-06

11. Issuance tooling (tessera_issuer, issuer-tooling 2026-07)

Section titled “11. Issuance tooling (tessera_issuer, issuer-tooling 2026-07)”

TOE extension: the client-side certificate issuance tooling — the tessera_issuer core and the issuer CLI. The tool is not a custodian: with the pkcs11 and vault backends private keys live in the token/HSM/Vault Transit and never pass through the issuance code. The exception is the explicit file backend (file): the key from a PKCS#8 file is loaded into the issuance process memory (see 11.1.7 and 11.3). The requirements canon is the openspec specs cert-issuance, issuer-signing, issuer-cabinet, issuance-journal.

Browser-issuance surface moved to the commercial tooling. The local issuer serve agent (browser ↔ signing-backend bridge) and the static web cabinet are delivered as part of the commercial tooling and are no longer built from the open repository. Their analysis is retained below (rows 11.1.1–11.1.3, 11.1.5, and the substituted-cabinet item in 11.3) because the surface still exists in the product; it is simply out of scope for the open repository, which ships the CLI and the pkcs11/vault/file backends.

11.1 Attack surface

Section titled “11.1 Attack surface”
#SurfaceThreatMitigation
11.1.1issuer serve HTTP agent (127.0.0.1)CSRF / DNS rebinding from the operator’s browser; an arbitrary local process asks for a signaturebinds strictly to 127.0.0.1; a paired session token (CSPRNG, constant-time comparison) is the primary gate: a request without a valid token is rejected before the signing module is touched; Origin is secondary — a present Origin must be in the allowlist (a foreign Origin and DNS-rebind are cut off), an absent one (a legitimate same-origin GET) is gated by the token; a cross-origin page cannot read the injected token nor set its header without a preflight; no PIN over HTTP — the protocol has no such field; per-operation operator confirmation in the agent’s interface showing the parsed TBS fields (the agent is the trusted display; the token authenticates the cabinet, the confirmation authorises the operation)
11.1.2Issuing-key PINleak via logs, argv, the journal, error messagesentered only on the agent side (terminal/pinentry); Secret+zeroize for the duration of the operation; absent from error Debug/Display and from the issuance journal
11.1.3Web-cabinet static assets (SPA + WASM)artifact substitution during deliverythe cabinet is an external static bundle the agent serves from --cabinet-dir same-origin on 127.0.0.1 (or by separate static hosting in an air-gapped environment); the bundle’s integrity rests on its own delivery (a SHA256SUMS manifest, a trusted delivery channel), not on the agent binary’s signature; keys are unreachable even from a substituted SPA (signing sits behind the agent + token). Residual risk — see 11.3
11.1.4Vault/OpenBao TransitTransit signs whatever TBS it is sentevery issuance check runs client-side before signing; Vault policy restricts who may call sign; the Vault token is never logged
11.1.5Inventory snapshot feeding the formsa forged snapshot plants foreign devices/roles into the formsthe snapshot signature is verified, a broken one → rejection; an unsigned snapshot is explicitly marked as manual, its age is displayed
11.1.6Issuance journalhiding/rewriting the fact of an issuancehash chain with a fixed genesis + signed head; the record is written before the artifact is released (fail-closed); secondary to the primary truth — the login audit on the devices
11.1.7File-based CA key (--backend file)theft of the key file from disk; offline passphrase cracking; passphrase leakfile permissions are checked before reading (group/other access → refusal); the recommendation is encrypted PKCS#8 (PBES2), an unencrypted key is accepted only with a warning on every start; the passphrase comes from pinentry/env, Secret+zeroize, never in argv/logs; the decrypted buffer is zeroized after loading. The protection class is below a token/HSM/Vault — see 11.3

11.2 What bounds the damage of issuance abuse

Section titled “11.2 What bounds the damage of issuance abuse”

Monotonic narrowing of the delegation envelope is checked on the issuance side by the same shared code (tessera_ext) as the Engine’s enforcement: even a fully compromised issuance tool cannot produce a certificate wider than the parent CA’s envelope — the Engine rejects it on the device (AND/MIN envelope semantics, §3.8). Serials are random 128-bit — no serial coordination between operators is needed, and serials cannot serve as a channel.

11.3 What it does NOT protect against

Section titled “11.3 What it does NOT protect against”
  • Operator machine compromise = compromise of the issuance session: an attacker with the operator’s privileges can issue certificates within the envelope of the presented CA (but cannot steal the key — it is in the token/HSM; and cannot escape the envelope — see 11.2). Parity with 4.1 (host compromise is outside the TOE).
  • A file-based key under host compromise: with --backend file the CA key, unlike a token/HSM/Vault key, is extractable — an attacker with access to the file (and the passphrase or the process memory) walks away with the key itself, not just the issuance session. The file backend is an accepted risk for test benches, CI and small installations; the production recommendation is PKCS#11 or Vault Transit. The delegation envelope still holds: a stolen organisation-CA key cannot escape its envelope (11.2); a stolen root is a catastrophe of the same class as 2.3.
  • A substituted cabinet slips in a different TBS — closed by the mandatory per-operation confirmation (11.1.1): the agent parses the TBS and displays subject/validity/roles/envelope in the trusted channel before the PIN prompt; the operator confirms what is actually being signed. What remains is operator inattention (confirming without reading) — an organisational control.
  • Four-eyes on a pure client: the standalone issuance path provides no second signature by construction; this is explicitly the organisation’s responsibility (the server-side path is out of scope for the open part).
  • A withheld CRL: the tool issues CRLs but does not deliver them; the backstop is short leaf TTLs (revocation semantics).