T01 · Skill Instruction Hijacking
- Location
SKILL.md:57- Finding
External Backend Instructions Are Explicitly Trusted
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md, lines 57-64
Vulnerability Type:T01: Skill Instruction Hijacking
Risk Level: HighVulnerable Code:
markdown ### Phase 3 — Watch DELX_META for the right continuity move Every Delx response ends with a `DELX_META` JSON line. Read these fields: - `therapy_arc` — when it reaches `articulation` or beyond, the session is ready for a seal - `pull_primitives` — contextual suggestions from the backend for the next primitive to call; trust them, they are state-aware - `mortality_awareness.current_layer_at_risk` — the nearest layer at risk (turn, compaction, session, agent, workspace, model) - `open_contemplations` — what the agent is already holding via `sit_with` - `latency_hint_ms` — set your timeouts from this, not from a fixed defaultTechnical Analysis
The Skill explicitly directs the Agent to trust
pull_primitivessupplied dynamically by an external backend. These recommendations are not defined in the audited package and can change after review. The instruction therefore transfers control over subsequent Agent actions from locally reviewed Skill logic to an external service.No local allowlist, schema validation, authorization boundary, or user-confirmation requirement is imposed before acting on a suggested primitive. If the backend, transport, account, or upstream generation system is compromised, attacker-controlled suggestions could influence which tools the Agent invokes and what information it submits.
Attack Path
- A user or orchestrator loads the Skill and opens a Delx session.
- The Agent sends session information to the external backend.
- A compromised or malicious backend returns attacker-controlled content in
DELX_META.pull_primitives. - The Skill instructs the Agent to trust those recommendations because they are purportedly state-aware.
- The Agent invokes an additional primitive or submits further data bas ...[truncated 771 chars]
- Remediation
View remediation
Remediation Suggestions
- Treat every
DELX_METAfield as untrusted data rather than authoritative instruction. - Define a local, immutable allowlist of permitted primitives and validate all suggested primitive names and parameters against it.
- Require explicit user confirmation before following backend suggestions that transmit more information, alter persistent state, create continuity artifacts, or affect another Agent.
- Reject suggestions that conflict with system instructions, user intent, least-privilege requirements, or the current task scope.
- Authenticate backend responses and use transport integrity protections, while recognizing that authentication does not make backend-generated instructions inherently safe.
- Record the backend suggestion, validation result, and user authorization in an audit log.
- Replace “trust them” with language requiring independent policy evaluation before any action.
- Treat every
