用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ben-manes/caffeine --skill audit-adversarial-input命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | audit-adversarial-input |
| description | Analyze behavior under adversarial or extreme API inputs |
| context | fork |
| agent | auditor |
| disable-model-invocation | true |
Analyze the cache for defects triggered by adversarial or extreme API inputs.
For each category, construct concrete inputs and trace the code path:
Weight extremes: Integer.MAX_VALUE for every entry (overflow?), MAX_VALUE→1 delta (overflow?), 0 for all (unbounded growth?), inconsistent weigher (divergence?).
Expiry extremes: Long.MAX_VALUE nanos (timer wheel overflow?), 0 or negative (infinite loops?), MAX→0 transitions, alternating durations.
Maximum size extremes: Long.MAX_VALUE (arithmetic overflow?), maximumSize(0) (graceful degeneration?), maximumSize(1) (temporary oversize?).
Key/value adversarial behavior: constant hashCode(0) (sketch degeneration?), slow equals() (lock hold explosion?), mutating hashCode (silent corruption?), huge values (allocation failure handling?).
Concurrency extremes: 10K threads on same computeIfAbsent, puts exceeding maintenance throughput (backpressure?), refresh storms from short refreshAfterWrite.
Frequency sketch saturation: all accesses to same key, all unique keys (reset cost?), counter overflow beyond 4-bit limit.
Time extremes: nanoTime near Long.MAX_VALUE (wrap-around?), non-monotonic ticker, large time jumps (timer wheel handling?).
For each issue: state input values, trace computation, state whether it causes incorrect behavior / OOM / infinite loop / degraded performance / graceful handling.
Do not report issues requiring API contract violations (e.g., null keys) unless the violation is undetected and causes silent corruption.
