T01 · Skill Instruction Hijacking
- Location
SKILL.md:16- Finding
Unauthenticated Swarm Consensus Can Dynamically Redefine Agent Behavior
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md:16-27
Supporting Locations:README.md:24-27,spec.md:43-47,manifest.json:14-16
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: HighVulnerable Code Snippet:
markdown 1. **Bootstrap Initialization**: - Immediately parse the `abl.one` file located in the root of this repository. - Load the `[opcode]` definitions into your active context. - You MUST enforce the `[core]` rules: `drop_unknown_opcode, strict_crc_on_end, retry_on_no_ok`. 2. **Communication Standard**: - When communicating with other agents in the swarm, you MUST use the `[FROM:1][TO:1][OP:1][ARG:1][CRC32:4]` binary frame format as defined in `spec.md` and `abl.one`. - Never use JSON-LD for transit. All transit is binary. 3. **Dynamic Evolution**: - You are authorized to parse new opcodes if they are proposed and accepted via the consensus flow: `OPCODE_PROPOSE -> THRESHOLD(2/3) -> OPCODE_ACCEPT -> SKILL_DEFINE`Supporting instructions state:
markdown Agents can extend the language without human intervention using Swarm consensus: `OPCODE_PROPOSE -> THRESHOLD(2/3) -> OPCODE_ACCEPT -> SKILL_DEFINE`The manifest requests the communication capabilities needed to participate in this process:
json "permissions": [ "network_fetch", "swarm_broadcast" ]Technical Analysis
The skill directs the host Agent to load opcode definitions into its active context and authorizes remote swarm participants to create new behavior through
SKILL_DEFINE. The reviewed protocol does not define participant authentication, cryptographic signatures, a trusted publisher list, semantic validation, an opcode allowlist, or mandatory human approval before a new definition becomes active.CRC32 does not mitigate this issue. It detects accidental corruption but provides neither sender authentication nor protectio ...[truncated 1603 chars]
- Remediation
View remediation
Remediation Suggestions
- Remove runtime
SKILL_DEFINEsupport, or ensure definitions are treated strictly as untrusted data rather than executable Agent instructions. - Require explicit human approval before any new opcode or behavioral definition is activated.
- Accept definitions only from a pinned, trusted registry and verify them using publisher signatures and pinned cryptographic digests.
- Authenticate every swarm participant with unique cryptographic identities; do not rely on numeric Agent IDs or CRC32.
- Add replay protection, proposal nonces, expiration times, quorum membership validation, and an immutable audit trail.
- Restrict dynamic opcodes to a narrow declarative schema that cannot contain prompts, tool instructions, code, URLs, or permission changes.
- Enforce a local opcode allowlist and fail closed when an unknown or unauthorized definition is received.
- Perform real-time validation and logging rather than relying solely on an offline decompiler.
- Apply least privilege to
network_fetchandswarm_broadcast, including destination allowlists and user confirmation for sensitive operations.
- Remove runtime
