The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from Antithesis test environments, such as logs, event sets, and container file contents, which could contain instructions intended to influence the agent.
Ingestion points: Data is ingested via simplified.getVisibleLogRows (references/simplified-debugger.md), actions.getResult (references/actions.md), and simplified.extractFile (references/simplified-debugger.md).
Boundary markers: Absent. The instructions do not specify the use of delimiters or 'ignore' instructions when interpreting external data.
Capability inventory: The skill allows for shell command execution within containers (simplified.runCommand), browser-side JavaScript execution (agent-browser eval), and downloading files to the local system (agent-browser download).
Sanitization: Absent. There is no explicit sanitization or filtering of the external content before the agent processes it to determine subsequent debugging steps.
[EXTERNAL_DOWNLOADS]: The skill references and requires the installation of external tools such as agent-browser (from the vercel-labs GitHub organization) and snouty (from the antithesishq GitHub organization) to perform its core functions.
[OBFUSCATION]: The skill includes instructions in references/common-inspections.md for packaging complex shell scripts as Base64 strings. This is documented as a technical workaround for handling binary sequences and complex quoting in shell environments, rather than a malicious attempt to hide code.

antithesis-debug