web-perf
Warn
Audited by Gen Agent Trust Hub on Oct 23, 2026
Risk Level: MEDIUMMETADATA_POISONINGINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [METADATA_POISONING]: The skill's YAML frontmatter identifies the author as "cloudflare" and the source as "cloudflare/skills", and includes a "verified" flag and a provenance hash. However, the actual provided author context for this skill is "oimiragieo". This discrepancy represents a deceptive use of metadata to misrepresent the skill's origin and authority, potentially leading users and agents to trust its instructions more than they should.
- [INDIRECT_PROMPT_INJECTION]: The audit workflow requires the agent to run performance auditing tools on user-provided target URLs and then analyze the resulting JSON reports. This ingestion of untrusted external data, combined with the agent's broad tool permissions (Bash, Write, Edit), creates a surface where a malicious website could inject instructions into the audit output to steer the agent's subsequent actions.
- Ingestion points: The skill reads and analyzes
lighthouse-report.jsongenerated from an external URL as part of the Phase 1 audit workflow. - Boundary markers: No specific delimiters or warnings are provided to the agent to treat the report content as untrusted or potentially adversarial data.
- Capability inventory: The skill uses
Bash,Write,Edit,Read,Glob, andGreptools, providing a wide range of actions that could be abused by injected instructions. - Sanitization: There is no evidence of sanitization or filtering of the performance report data before it is interpreted by the agent.
- [EXTERNAL_DOWNLOADS]: The skill's instructions include running
npx lighthouse, which involves downloading and executing the Lighthouse package. While this is a well-known tool from a reputable source, it remains an external runtime dependency that the agent is instructed to execute. - [COMMAND_EXECUTION]: The skill explicitly instructs the agent to use the
Bashtool to executenpxcommands for performance auditing, which involves running shell commands with arguments derived from potentially untrusted inputs (the target URL).
Audit Metadata