Security audit
Security checks for vulnerabilities and agentic risk
This is a Markdown-only reasoning guide that asks the agent to verify current claims and avoid over-extrapolation; its broad activation wording is worth noticing but not a security concern by itself.
Install this if you want the assistant to be more careful about current claims, market-gap assertions, and sparse documents. Expect it to trigger broadly in strategy or research conversations and to encourage live searches before claims about what exists now.
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.
2. **Apply them silently** during analysis — don't lecture the user about the principles
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.
These principles are reference material. The SKILL.md handles the high-level activation. When you (the AI applying this skill) face a turn that triggers one or more principles:
1. Identify which principles apply (use [triggers.md](triggers.md) for quick lookup)
2. Apply the discipline silently — don't lecture the user
3. Surface an epistemic note only when bypassing a principle would mislead
When unsure about how to apply a principle to a specific case, see [anti-patterns-catalog.md](anti-patterns-catalog.md) for sanitized concrete cases.
The phrase 'Activate when user asks how to discuss product/strategy questions, requests analysis of unfamiliar markets, or when sparse documentation might tempt extrapolation' mixes subjective conditions like 'might tempt extrapolation' with broad domains like product and strategy. This leaves unclear when the skill should activate versus remain inactive.
The README says the skill activates automatically when users discuss strategy, market, or product questions, or share sparse documentation, with no explicit invocation needed. These are broad, everyday discussion categories, and the file does not provide a precise trigger list or exclusion criteria at the activation point, which could cause unintended invocation.
The description says to activate when users ask about product/strategy questions, unfamiliar markets, or whenever sparse documentation might tempt extrapolation. These conditions are conceptually broad and lack clear boundaries or exclusions, increasing the risk of unintended invocation in many normal analytical discussions.
Items such as evaluating multiple directions, sparse documentation being in play, sharp pushback, and very short replies are common conversational patterns that may occur across many unrelated tasks. Because the file does not provide negative examples or tighter activation constraints here, the skill could activate too broadly.
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.
**Principle violated**: 6 (no premature frame-merging)
**Pattern**: User has Project A (work track) and Project B (side track). AI uses A as framing for B, or vice versa, without checking which track the topic belongs to.
**Concrete shape**: User asks about Project A's architecture. AI ends analysis with "and these primitives also give you tools for Project B!" when Project B wasn't part of the question.
This markdown file defines activation triggers, so SQP-1 applies. The section uses very common phrases like "What do you think?" and "What would you do?" as triggers for a specific principle, but it does not provide constraints or negative examples to distinguish ordinary conversational use from intended activation cases, increasing the risk of unintended invocation.
Phrases such as "Evaluate this for me" and "What are the pros and cons of X?" are broad and commonly used across many contexts. The file does not define boundaries for what kinds of artifacts or evaluation contexts should activate this sub-trigger, which could cause the principle to fire too broadly.
The short-reply trigger treats common acknowledgments like 'ok', 'yes', and very short replies as implicit instructions to continue without clarification. That can cause the agent to act on ambiguous user input, increasing the risk of unwanted disclosure, misexecution, or carrying forward an incorrect frame when the user did not actually consent to the inferred continuation.
The trigger explicitly treats Chinese phrases like "嗯", "继续", and "更多" as activation inputs alongside English terms, but there is no statement that language handling is configurable or user-selected. This can be read as imposing specific locale assumptions rather than offering language/locale choice.
No suspicious patterns detected.