brida-reflex-execution-failure-classification
Pass
Audited by Gen Agent Trust Hub on Sep 21, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill fetches live availability claims and documentation from the vendor's website (e.g.,
/reflex/llms.txt,/reflex/agent.md,/reflex/agent.json). These are standard operational requests to the skill author's infrastructure. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and classify unstructured data from build outputs and execution states, which may contain malicious instructions if the source is compromised.
- Ingestion points: Processes 'bounded execution/tool/build output' and environment events as described in
SKILL.mdandplaybook.md. - Boundary markers: Includes a dedicated 'Authority' section in
SKILL.mdandplaybook.mdspecifying that the classification is recommendation-only and prohibits protected actions like payments or merges. - Capability inventory: Integration via MCP, SDK, or REST for classification logic; no direct shell execution of untrusted data is requested.
- Sanitization: Requires normalization of bounded state and restricts processing to declared failure branches, with unknown results going to a safe fallback.
Audit Metadata