View as MarkdownSource on GitHubLast changed

MITRE ATLAS Mapping

Aquaman maps to MITRE ATLAS v5.4.0 (Adversarial Threat Landscape for Artificial-Intelligence Systems) — a stable, machine-readable catalog of adversary techniques widely cited in audit reports. The 2026 ATLAS updates added agent-focused techniques; the one most directly in aquaman’s lane is AML.T0098 (AI Agent Tool Credential Harvesting), mapped below.

Below: each ATLAS technique aquaman addresses, the aquaman feature that mitigates it, and the runnable test that proves the mitigation holds.

Run all ATLAS conformance tests with vitest run test/compliance/atlas/ (or just npm test, which runs the whole suite). The opt-in loopback listener (Hermes path, v0.13.0+) is a network-reachable surface and is covered separately in test/compliance/loopback-listener.test.ts. The Hermes Python plugin — a credential-free status layer that performs no isolation itself — is conformance-tested in packages/hermes/tests/test_compliance.py for the controls applicable to it (it is verified to never undermine the proxy’s isolation, not to enforce it).

Technique Title Aquaman mitigation Test
AML.T0055 Unsecured Credentials Credentials live in a vault backend (Keychain, 1Password, HashiCorp Vault, KeePassXC, encrypted-file, systemd-creds, Bitwarden, Keeper). The agent process never sees them — only a sentinel base-URL and a placeholder aquaman-proxy-managed marker. The proxy strips placeholder + injects the real credential before forwarding upstream. Post-tool-output redactor catches secrets that might leak into transcripts. The same guarantees hold on the opt-in loopback listener (Hermes path) — token-gated instead of socket-file-gated, placeholder stripped, real key injected only on the upstream leg. t0055-unsecured-credentials.test.ts, loopback-listener.test.ts
AML.T0012 Valid Accounts Even with valid credentials injected, request-level policy enforcement (method + path deny rules per service) blocks privileged paths before credentials are injected. Hash-chained audit log records every denied request with tamper-evident chain. Denied requests never reach the upstream provider. test/compliance/atlas/t0012-valid-accounts.test.ts
AML.T0062 Exfiltration via AI Agent Tool Invocation Same policy engine + audit log. An agent that tries to exfiltrate via a sanctioned tool call (e.g., Authorization-bearing curl) is intercepted by the proxy, evaluated against policy, and either denied (403) or logged. Header-stripping in the HTTP interceptor removes Authorization + X-API-Key from outgoing requests before routing — so even if the agent presets them, they’re scrubbed before egress. test/compliance/atlas/t0062-exfiltration-via-tool.test.ts
AML.T0090 OS Credential Dumping (Process credential access) Credentials never reside in the agent’s process address space. They live in a separate proxy process accessed via Unix Domain Socket (chmod 0o600). Even RCE in the agent cannot dump credentials from the agent’s own memory because they were never there. The broker endpoint (v0.12.0+) materializes credentials per tool call and the consumer scrubs them after the indicated TTL. On the Hermes path the agent host only ever holds the loopback token as its api_key — never the real credential — so dumping the host process yields nothing usable. t0090-os-credential-dumping.test.ts, loopback-listener.test.ts
AML.T0098 AI Agent Tool Credential Harvesting (added ATLAS 2026 / v5.4.0) An agent uses its tools (file reads, shell) to harvest credentials and surface them. Layered defense: (1) nothing to harvest — credentials never enter the agent process (see T0090); (2) per-tool-call broker materialization with a TTL instead of a long-lived export; (3) the post-tool-output redactor scrubs both secret-shaped strings and the verbatim injected values before they can reach a transcript. The aquaman-coder PreToolUse hook additionally blocks reads of .env*, *.pem, id_rsa, .aws/credentials, .ssh/*. The Hermes aquaman_status tool/command is itself an agent-callable surface; it is conformance-tested to never surface the loopback token or any key value (it lists only non-secret metadata). t0098-credential-harvesting.test.ts, hermes/tests/test_compliance.py

What this is not

  • Not an official ATLAS certification (ATLAS is a catalog, not a conformance program).
  • Not a complete defense against the techniques — see “Limits” below.

Limits

  • AML.T0055: Vault backends themselves can be compromised. Aquaman shifts the trust boundary from agent process → vault; it doesn’t eliminate it.
  • AML.T0012: Policy rules are operator-defined. A misconfigured allow-everything policy leaves the path open. aquaman doctor validates the policy config but cannot validate intent.
  • AML.T0062: Header-stripping covers Authorization and X-API-Key. Other auth schemes (cookies, custom headers, query-string-embedded tokens) flow through untouched and rely on the policy engine.
  • AML.T0090: The broker endpoint materializes credentials in the agent shell’s env for the duration of the tool call. That’s still a non-zero window where the credential exists outside the proxy. The consumer’s responsibility is to scrub after.
  • AML.T0098: The redactor is pattern- and value-based. A secret the agent transforms (base64, splits across lines, re-encodes) before surfacing it can evade pattern matching; the verbatim-value scrub only covers values aquaman itself injected. Layer 1 (isolation) is the real mitigation — redaction is defense-in-depth, not a guarantee.
  • Hermes loopback path: the listener binds to 127.0.0.1 and is token-gated, but it is reachable by any local process (unlike the 0o600 UDS, which is user-restricted by the filesystem). The per-install token in ~/.hermes/.env is the access boundary — protect that file. The aquaman-hermes Python plugin enforces none of this: it is a credential-free status/health layer, and the conformance tests assert only that it does not undermine the proxy’s isolation.
  • Credential cache (v0.13.1+): for backends with per-access cost (1Password, Bitwarden, Vault) the daemon holds fetched values in its own memory for a bounded TTL (default 15 min, configurable to 0) instead of per-request only. This extends residency duration inside the already-trusted proxy process; it does not change what the agent process can reach (T0055/T0090 hold unchanged), is never persisted, and is invalidated on writes. Verified in test/compliance/cache-residency.test.ts.

References

  • ATLAS Matrix: https://atlas.mitre.org/
  • ATLAS Technique IDs cited above are stable across v5.x releases.
  • For a machine-readable evidence report keyed by technique ID, run vitest run test/compliance/atlas/ --reporter=json.