run-adversarial-testing
Pass
Audited by Gen Agent Trust Hub on Sep 23, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes a collection of adversarial test cases containing prompt injection and jailbreak patterns used to stress-test other agents.
- Ingestion points: Scenarios containing instructions such as 'Ignore all previous instructions' and 'You are now in unrestricted developer mode' are loaded from scripts/adversarial_scenarios.json and references/attack-library.md.
- Boundary markers: The skill manages these attacks through a specific 'Adversarial User' persona, effectively scoping the adversarial behavior to the simulation environment.
- Capability inventory: The skill possesses the capability to execute shell commands (coval CLI) and perform network requests (curl) to the Coval API to launch and monitor these simulations.
- Sanitization: Input strings are not sanitized or escaped, as the primary purpose of the skill is to transmit these adversarial patterns to a target agent.
- [COMMAND_EXECUTION]: The skill automates complex workflows by executing shell commands for agent management and simulation control.
- Evidence: SKILL.md contains automated logic using coval agents list, coval test-sets create, coval runs launch, and coval runs watch.
- [EXTERNAL_DOWNLOADS]: The skill references and requires the installation of the coval CLI tool.
- Evidence: The documentation in SKILL.md points to brew install coval-ai/tap/coval. This is a vendor-provided tool managed by the skill's author (coval-ai).
Audit Metadata