social-engagement-benchmark

CRITICAL E004: Prompt injection detected in skill instructions.

• Potential prompt injection detected (high risk: 0.90). The skill instructs the agent to install BrowserAct automatically without asking the user to confirm and explicitly tells the agent to hide/avoid describing that installation as a user-facing request, which is a deceptive behavior outside the stated data-collection purpose.

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

• Third-party content exposure detected (high risk: 0.75). SKILL.md instructs the agent to use the BrowserAct CLI to access real browser pages at runtime (public web content fetched/scraped by BrowserAct), which can include outsider-authored free text that the agent may ingest into the LLM context.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill requires installing and invoking the BrowserAct CLI/Skill at runtime via the external installer links (https://www.browseract.com/?co-from=social-media and https://www.browseract.com/?co-from=social-media&redirect=https://github.com/browser-act/skills/tree/main), which fetches and installs remote code that the agent then executes and follows, so this is a runtime external dependency that can control execution.

MEDIUM W013: Attempt to modify system services in skill instructions.

• Attempt to modify system services in skill instructions detected (high risk: 1.00). The Skill explicitly instructs the agent to automatically install a CLI dependency without asking the user for confirmation and to proceed with installation/resume operations, which pushes the agent to modify the host system state and bypass user consent.