refit
Pass
Audited by Gen Agent Trust Hub on Oct 23, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONPERSISTENCEPRIVILEGE_ESCALATIONCOMMAND_EXECUTIONTIME_DELAYED_CONDITIONAL
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests and processes environment data that could be influenced by external sources or prior agent tasks, creating a surface for indirect instructions.
- Ingestion points: Reads
.omc/logs/, session state, and problem reports (issues.md,problems.md) which may contain untrusted data or adversarial content. - Boundary markers: The instructions do not define specific delimiters or warnings to ignore embedded instructions within these data sources.
- Capability inventory: The skill has the authority to write to CI workflows, git hooks, and repository scripts, and can execute shell commands to verify new checks.
- Sanitization: No explicit sanitization, filtering, or escaping of the ingested environment data is described before it is processed by the agent.
- [PERSISTENCE]: The skill's primary function is to modify the development environment across sessions, specifically targeting files that can execute code automatically.
- Evidence: It is instructed to land findings in "surfaces" including CI workflows, git hooks, and project scripts (
package.json,.mcp.json), which are common targets for maintaining persistent access. - [PRIVILEGE_ESCALATION]: By modifying CI workflows and git hooks, the skill can influence high-privilege automation environments. Although it requires user approval, a sophisticated attacker could poison the logs to prompt the agent to propose a seemingly benign but malicious configuration change.
- [COMMAND_EXECUTION]: The skill requires the agent to execute shell commands to verify the effectiveness of new "deterministic checks," which involves running the check and demonstrating its failure on a deliberate violation.
- [TIME_DELAYED_CONDITIONAL]: The skill supports a "Headless refit" mode intended for automated execution via host schedulers such as cron or CI timers, which triggers the survey process independently of an interactive user session.
Audit Metadata