autoresearch

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 1.00). The skill's ground phase explicitly WebFetches and saves arXiv HTML to .autoresearch/refs (phases/ground.md) and the experiment phase allows adding arbitrary external repos/datasets via git submodule add (phases/experiment.md), which the agent is required to read and use to drive experiments and decisions, exposing it to untrusted public content that could carry indirect prompt injections.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.80). The skill's ground phase instructs the agent to WebFetch and save full arXiv HTML (e.g., https://arxiv.org/html/) at runtime and use those files as full-text context injected into the agent's prompts, so remote content can directly control the agent's instructions.