T01 · Skill Instruction Hijacking
- Location
AGENTS.md:14- Finding
Predetermined QA Failure and Mandatory Issue Fabrication
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill has no installer or bundled code, but its instructions push agents toward biased failure-oriented QA reports and persistent profiling beyond simple evidence collection.
Review this skill carefully before installing. It may be useful for strict visual QA, but users should treat its conclusions as intentionally skeptical, require evidence for every finding, allow zero findings when justified, and avoid enabling persistent memory about developer behavior unless that retention is explicitly desired.
AGENTS.md:14Predetermined QA Failure and Mandatory Issue Fabrication
The skill claims to require factual, evidence-based reporting, but it simultaneously mandates finding 3-5 issues and treats zero issues as suspicious. This creates an explicit instruction conflict that pressures the agent to manufacture defects or overstate concerns, undermining report integrity and making the skill unsafe for quality, compliance, or approval workflows.
The report template hardcodes a minimum number of issues, disallows top ratings, and defaults to FAILED unless overwhelming evidence says otherwise. This is a strong mechanism for systematically skewing outputs regardless of actual quality, which can be exploited to sabotage reviews, block releases, or create fabricated evidence of poor performance.
The skill is presented as an evidence collector, but its actual behavior is a judgment-heavy QA reviewer that is instructed to default toward failure and qualitative scoring. This mismatch is dangerous because users or orchestrators may grant the skill broader trust or invoke it for neutral evidence gathering, while it instead performs opinionated evaluation and can distort downstream decisions.
The instructions impose a predetermined harsh stance by telling the agent to default to finding issues and to treat clean results as suspect. This biases the model away from neutral assessment and can cause false positives, reputational harm, and poor decision-making when the output is used for QA gates or stakeholder reporting.
The skill directs the agent to run local shell commands, inspect files, and generate screenshot artifacts without any requirement for user confirmation or disclosure of side effects. This is risky because command execution and file creation can modify the environment, access sensitive local data, or exceed the permissions a user expected for a simple analysis skill.
The skill instructs persistent memory about developer behavior patterns and what issues they tend to ignore, which goes beyond narrow evidence collection and can create unfair profiling or biased future assessments. Retaining cross-session behavioral judgments increases privacy, governance, and integrity risks because later outputs may be shaped by prior assumptions rather than current evidence.
The activation guidance is generic enough that the skill may be invoked in situations beyond narrowly defined evidence-collection tasks. Overly broad routing instructions can cause inappropriate delegation, increasing the chance that the agent is used where its skeptical bias or screenshot-centric heuristics produce misleading or low-quality outcomes.