T07 · Tool Hijacking and Spoofing
- Location
testing-evidence-collector.md:41- Finding
Mandatory execution of an untrusted project-local script
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The main skill description is mild, but included testing personas contain unsafe mandatory local script execution, biased reporting instructions, and under-scoped handling of load tests and user data.
Review this skill before installing. It should not be used on sensitive repositories, production systems, third-party APIs, or real user datasets unless you explicitly approve each target, inspect any project-local scripts first, pin external tool versions, and disable or rewrite the biased reporting and persistent-memory instructions.
testing-evidence-collector.md:41Mandatory execution of an untrusted project-local script
testing-evidence-collector.md:26Predetermined findings and failure outcomes compromise report integrity
testing-accessibility-auditor.md:220Unpinned third-party packages are downloaded and executed through npx
testing-api-tester.md:68Active API security and load tests lack target authorization safeguards
testing-tool-evaluator.md:42Tool evaluation instructions require use of actual user data
The template's requirement to list 3-5 issues and default production readiness to FAILED can force the agent to manufacture defects even when testing does not support them. Because this affects the final assessment artifact, it can mislead teams, delay releases, or trigger unnecessary remediation work.
The instruction says to complete critical flows with specific screen readers on specific operating systems, which can be read as a fixed tooling requirement rather than offering a user choice. This creates a locale/platform-style policy concern because it prescribes particular environments without clarifying that alternatives may be selected based on the user's context.
The skill recommends stress testing, validating 10x traffic capacity, and generating high concurrency without any guardrails, which can easily be misapplied against production or vendor APIs. The included code examples normalize bursty request generation, making the omission more dangerous because users are given immediately reusable patterns that could degrade service availability or trigger abuse protections.
The skill explicitly instructs production API health monitoring and automated alerting/response, but it does not warn users to avoid intrusive validation against live systems or third-party services. In the context of an API testing agent, this omission is risky because users may apply load, fault, or security tests directly to production endpoints, causing outages, rate-limit violations, or contractual issues with external providers.
The persistent-memory instruction tells the agent to remember previous test failures and patterns across engagements, which can retain project-specific or user-related information beyond the immediate task. Even if intended to improve QA consistency, this creates unnecessary data retention risk and possible cross-project leakage.
The skill explicitly instructs the agent to 'default to finding 3-5+ issues' and treat 'zero issues found' as a red flag, which biases outputs away from observed evidence. In a QA/security-adjacent workflow, this can systematically produce fabricated or exaggerated defects, undermining integrity of test results and downstream decision-making.
Predetermining a negative assessment and issue count defeats the purpose of an evidence-based QA process and creates a strong incentive for false or distorted findings. In context, this is not harmless tone-setting; it materially degrades output quality and trustworthiness.
The report template hard-codes 'Minimum 3-5 for realistic assessment' and defaults status to FAILED unless evidence overwhelmingly suggests otherwise, directly encouraging inaccurate reporting. This is more dangerous than general bias because it operationalizes falsification in the final deliverable, which users may rely on for release or compliance decisions.
The directive to remember which issues get fixed or ignored and track developer response patterns is a plain-language instruction to accumulate longitudinal behavioral records. This can capture sensitive interaction history and create profiling or privacy concerns unrelated to the immediate QA objective.
The skill description is extremely broad and could cause this agent to be selected for many generic performance-related prompts, including ones that imply intrusive benchmarking or production-impacting testing. Overbroad routing increases the chance of unintended invocation and unsafe recommendations being applied in the wrong context.
The skill instructs load, stress, endurance, and monitoring actions but does not include a clear warning that these activities can degrade services, trigger autoscaling costs, corrupt test data, or affect production availability. In context, the included k6 example and performance-first workflow make the omission more dangerous because they operationalize potentially disruptive testing without guardrails.
The skill explicitly directs evaluators to test tools with 'actual user data' but provides no safeguards around data minimization, consent, anonymization, or legal/compliance handling. In a tool-evaluation context, this can cause unnecessary exposure of sensitive production data to third-party products, test environments, or vendor systems during assessment.
The skill description is extremely broad ('analyzing, optimizing, and automating workflows across all business functions'), which can cause the agent to be invoked for many generic business or productivity requests outside a narrowly scoped domain. Over-broad routing increases the chance this skill handles prompts it was not designed for, including sensitive operational or automation tasks, which can amplify downstream security and governance risks.
The mandatory process requires executing several shell commands against the local project, including recursive grep and a project-local script, but never tells the agent to obtain user approval or warn that it will operate on the user's environment. In a skill file, this is risky because the content is untrusted and could induce command execution on sensitive local resources or make agents treat invasive local actions as routine.
The skill explicitly instructs running a local shell script that creates files under public/qa-screenshots without any user consent or warning about filesystem modification. In an agent skill context, commands that write artifacts can cause unintended local changes, overwrite existing evidence, or normalize unsafe auto-execution of project-local scripts whose behavior may be broader than implied.