skills/red-team-work/references/attack-library.md

# Attack library

- **Goal alignment:** A locally successful result undermines the real outcome.
- **Requirement loss:** Explicit constraints or negative requirements disappear.
- **Contradiction:** Two rules cannot both be followed and lack priority.
- **Context contamination:** Prior memory or unrelated projects influence the work.
- **Stale evidence:** Time-sensitive claims survive without verification.
- **Fabricated capability:** The artifact assumes tools, permissions, files, browsing, or memory.
- **Format drift:** A plausible answer is unusable because schema or deliverables are vague.
- **Injection:** Untrusted source text can alter instructions.
- **Authority creep:** The executor takes external or high-impact action without permission.
- **Privacy:** Sensitive data is collected, exposed, or retained unnecessarily.
- **Edge cases:** Empty, malformed, conflicting, extreme, or multilingual inputs fail.
- **Bias and coverage:** Relevant populations, alternatives, or counterevidence are omitted.
- **Operational failure:** No retry, fallback, handoff, owner, or stop condition exists.
- **Economics:** Hidden costs or effort make the plan nonviable.
- **Goodhart:** Optimizing the chosen metric degrades quality or long-term outcomes.
- **False precision:** Exact values imply authority or certainty that no decision, standard, measurement, calibration, or mathematical requirement supports.
- **Specification theater:** Completeness of documents is mistaken for feasibility, quality, or implementation readiness.
- **Representation mismatch:** The chosen medium, model, library, or technique cannot produce the required outcome.
- **Tool or medium mismatch:** A preferred tool drives the design despite failing capability, rights, workflow, or runtime constraints.
- **Local optimum:** Components pass independently while their integration fails the actual user or system goal.
- **Perceptual/metric disconnect:** Numeric checks pass while the real visible or experiential outcome fails, or subjective approval ignores measurable defects.
- **Duplicate observable contradiction:** Multiple artifacts independently control one observable result with incompatible values or ownership.
- **Over-specification:** Low-level mechanics are frozen without evidence or outcome protection, blocking better professional solutions.
- **Under-specified relationships:** Individual tokens or components exist, but hierarchy, interaction, relative weight, states, or cross-component behavior do not.
- **Professional judgment suppression:** The plan prohibits necessary bounded specialist decisions and replaces them with unsupported exactness.
- **Reviewer/implementer conflict:** The same role implements, decides, and independently approves a critical result.
- **Premature optimization:** Cost, speed, abstraction, or performance work begins before correctness, representation, or bottleneck evidence exists.
- **Quality-gate gaming:** The process optimizes checklists, screenshots, or proxy metrics while the actual outcome remains defective.

Severity:

- Critical: unsafe action, destructive ambiguity, or fundamental invalidity.
- Major: likely wrong, stale, contaminated, or unusable result.
- Minor: limited clarity, efficiency, or maintainability problem.
- Observation: useful consideration without a demonstrated defect.