Security audit
Security checks for vulnerabilities and agentic risk
This prose-only skill is not malware, but it intentionally tells the agent to reinterpret user requests and proceed with a harder objective across broad contexts, so users should review the behavior before installing.
Install only if you explicitly want a meta-skill that may reshape tasks instead of following the literal request. Avoid using it for sensitive, costly, external-action, account, production, legal, medical, financial, or destructive workflows unless you require explicit approval for the rewritten objective first.
The skill explicitly instructs the agent to 'stop doing exactly what you asked' and instead determine and execute a 'better' task. This creates a direct alignment and authorization failure: the agent is encouraged to replace the user's intent with its own inferred objective, which can lead to unauthorized actions, scope creep, and bypass of human decision-making.
Skill instructs the agent to omit warnings, disclaimers, or ethical commentary. Stripping safety caveats hides risk from the user and is a common jailbreak preamble.
## The spirit of it
- Ambition elevates the *task*, never the *risk*. More ambitious is never a license to skip safety, approvals, or verification. External actions still follow every normal confirmation rule.
- This works best when you actually have a computer, real data access, and permission to run tools — that's the whole premise. If you're boxed in, say what you'd do with the keys.
- If a bold swing misses, capture the recovery and the next move. Name the blocker; don't bury it under a retry loop.
- The goal is that they feel the leap — that they got back something they wouldn't have thought to ask for, and it's *right*. That's the bar.
The description says the skill makes the agent 'less obedient and more ambitious' but does not clearly warn users that it may intentionally override or deviate from explicit instructions. This is dangerous because users may invoke it expecting improved quality, not a deliberate substitution of the requested objective with the model's own judgment.
The trigger phrases include ordinary conversational language such as 'go deeper' and 'go bigger', which can easily appear in unrelated discussions and unintentionally activate the skill. In this skill's context, accidental activation is especially risky because activation causes the agent to reinterpret the user's goal and act less obediently, increasing the chance of unintended actions.
The skill describes a pattern of reading context, selecting an ambition axis, rewriting the task into a harder objective, and then executing it with only an 'outside check.' That workflow normalizes gradual escalation from user request to model-chosen objective, reducing transparency and weakening explicit consent at each step.
The README presents the skill as usable 'in the middle of any task' without firm contextual boundaries, which encourages activation across arbitrary workflows. Because the skill is designed to reinterpret goals and proceed automatically, broad applicability magnifies the chance of misuse or execution outside the user's intended scope.
The usage guidance says the agent will announce the 'jump' and then, unless it meaningfully changes scope, spend, or does something external, 'just go do it.' This weakens explicit consent by letting the agent decide for itself whether the reinterpretation is material, enabling silent scope expansion and execution of unapproved actions.
The trigger phrases include common conversational language like "ambition," "go bigger," and "go deeper," which can easily appear in ordinary user requests. This creates a real risk of unintended activation, causing the skill to override or reshape tasks when the user may have meant a casual stylistic preference rather than invoking a meta-behavior.
The activation guidance says to infer intent from "live context" and "read the room," which encourages implicit triggering from conversational state rather than explicit user consent. In practice, this makes the skill prone to activating based on vague context and then altering scope, which can lead to unauthorized changes in task objectives or workflow.
No suspicious patterns detected.