article-extractor
Pass
Audited by Gen Agent Trust Hub on Oct 26, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONPROMPT_INJECTIONCOMMAND_EXECUTIONMETADATA_POISONINGCREDENTIALS_UNSAFE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill implements a 'Self-Evolving' mechanism that encourages the agent to modify its own
SKILL.mdinstructions. Since the skill's primary function is to ingest untrusted data from the web (mql5.com), this creates a significant vulnerability surface where malicious content embedded in a scraped article could influence the agent to rewrite its persistent instructions with malicious content. - Ingestion points: HTML content and code blocks fetched from
mql5.comarticles and documentation pages. - Boundary markers: None identified; external content is parsed and converted directly into Markdown for storage and reference.
- Capability inventory: The skill uses high-privilege tools including
Bashfor network operations (curl) and local script execution within the working directory. - Sanitization: There is no documented validation or sanitization of ingested content before it is processed by the agent in a context where self-modification is encouraged.
- [PROMPT_INJECTION]: The
SKILL.mdfile contains a 'Self-Evolving Skill' section that explicitly instructs the agent to 'fix this file immediately' when encountering issues or workarounds. This meta-instruction effectively encourages the agent to override its static safety boundaries and persistent configuration, increasing the risk of state compromise through adversarial inputs. - [COMMAND_EXECUTION]: The skill makes extensive use of the
Bashtool to executecurlfor fetching data and local Python and shell scripts (e.g.,scripts/extract_all_python_docs.sh). This grants the agent broad execution capabilities over the local environment to facilitate scraping tasks. - [METADATA_POISONING]: The
references/evolution-log.mdcontains detailed narratives about previous 'Adversarial reviews' and 'Regression controls,' which could mislead users or automated scanners into believing the skill has undergone rigorous security validation. Additionally,references/troubleshooting.mdcontains a deceptive claim that 'No network tools [are] allowed,' which is contradicted by the skill's explicit use ofcurlandBashfor network access. - [CREDENTIALS_UNSAFE]: The
references/evolution-log.mdfile explicitly references a specific git commit (acc0cc3) where credentials were previously exposed in the repository history. While the credentials are not in the current active code, the metadata provides a direct pointer to sensitive information leaked in the version control history.
Audit Metadata