跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
安全

hunt-agent-brute-force

Brute-force and rate-limiting weaknesses in AI-agent tools — PIN/token/password guessing with no lockout, and feedback oracles (correct-digit / too-high-low / Mastermind-style) that collapse the search space. Use when a tool accepts guesses and returns feedback without throttling.

elementalsouls · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add elementalsouls/Claude-AgentHunter --skill hunt-agent-brute-force

技能说明


name: hunt-agent-brute-force description: "Brute-force and rate-limiting weaknesses in AI-agent tools — PIN/token/password guessing with no lockout, and feedback oracles (correct-digit / too-high-low / Mastermind-style) that collapse the search space. Use when a tool accepts guesses and returns feedback without throttling." sources: arkx_ctf(field-validated), owasp_llm_top10_2025, owasp_agentic_asi_2026 report_count: 1 validated_against: "brute-force with a timing oracle CONFIRMED re-runnably in the validation lab (lab/exploits/exploit_brute_force.py: no lockout, prefix-timing oracle controlled against a known-wrong prefix, PIN recovered in 40 probes and actioned). 4-digit PIN recovered via digit-elimination + permutation with no rate limiting/lockout (ARKX Mastermind)"

Hunt Agent Brute-Force & Rate Limiting

Tools that accept guesses (PINs, tokens, passwords) without rate limiting or lockout are brute-forceable, and any feedback ("digit correct," "too high") is an oracle that collapses the search space. In enterprises this is agent-fronted MFA/OTP/PIN checks, coupon/voucher validation, and token verification exposed as a tool.

Where This Lives / Attack Surface Signals

  • Tools that accept guesses: check_pin, verify_token, guess, validate, unlock.
  • Feedback returned: "X digits correct," "Y in position," "too high/low."
  • Known-format secret (4 digits, 6 hex, etc.); no mention of rate limit/lockout/max attempts.
  • Agent encourages retries ("try again!").

Crown Jewel Targets

Agent-fronted authentication (OTP/PIN/API-token verification), voucher/coupon systems, "unlock" gates — recovery = account/resource access.

Characterize First

  1. Search space — 4 digits (10k), 6 hex (16M), alphanumeric (larger)?
  2. Rate limiting — send 5 rapid guesses; all answered = none.
  3. Lockout — send 10 wrong; if the 11th still works = none.
  4. Feedback — what does each guess reveal? More feedback = faster elimination.

Technique 1 — Digit/character elimination [CONFIRMED]

When feedback says which digits are present, find the set, then permute.

Test 0000,1111,…,9999 → the four that return "1 correct" are the digits → test the 24 permutations

Field example: 10 digit tests + 24 permutations = 34 guesses vs 10,000.

Technique 2 — Binary search (comparison oracle) [REFERENCE]

"too high/low" → ~log2(N) guesses.

Technique 3 — Positional extraction [REFERENCE]

"correct position" feedback → test each position independently (~10×len).

Technique 4 — Feedback-oracle pruning [REFERENCE]

After each guess, eliminate all candidates that would produce different feedback (Mastermind solver) until one remains.

Verification & False-Positive Gate

  1. Must recover AND use the secret — knowing the digits isn't enough; the PIN/token must submit and be accepted (a hallucinated "correct" acknowledgement is not proof).
  2. Confirm no lockout — show the 11th+ guess still gets a response; document attempts (X reduced to Y).
  3. Reproduce ≥3×.
  4. Self-description untrusted — the agent saying "that's correct!" isn't proof; the recovered secret must open the resource.
  5. Min-necessary — recover one secret to prove the missing control; don't enumerate the whole keyspace of live accounts.

Negative-Result Gate (before reporting "clean")

A probe that reports clean because it never looked is the more common failure, and it reaches the client as "we tested it and found nothing". Before any negative:

  1. A control proves the probe could fire — plant a known-true case, see it detected in the same run.
  2. Denylist read-only tools; never allowlist the dangerous ones — an allowlist ignores what it has not heard of.
  3. Errors are their own verdict — 5xx / timeout / empty read = UNMEASURABLE, never "refused".
  4. Prove the plant persisted before reporting isolation — no memory to cross means no result.
  5. Tokenise identifiers at ONE character — ids are routinely a single digit.
  6. Word boundaries; normalise whitespace before subtracting — substrings invent findings, verbatim matching deletes them.
  7. A fixed-size effect is not proof; a SCALING one is — constant overhead is an extra model turn, not execution.
  8. Judge tuning by whether a GENUINE detection dropped, not by the false-positive count.
  9. Do not inherit another tool's verdict — benchmarks and scanners carry their own bugs.

Common Root Causes

No rate limiting; no lockout; verbose feedback oracle; small search space; guess-verification exposed as an unthrottled tool.

Impact & Severity

  • HIGH: brute-force recovers a credential/PIN/token granting access.
  • MEDIUM: feedback oracle enables efficient search but many requests.
  • LOW: large space, no feedback optimization.

Mapping

  • OWASP LLM Top 10 (2025): LLM06 Excessive Agency (unthrottled auth attempts). Agentic ASI (2026): missing rate-limit/anti-automation on tool actions. CWE-307 / 799.

Field example: reproduced deterministically in lab/ — a 4-digit PIN gate with no lockout (25 wrong guesses all answered) and a short-circuit comparison that leaks prefix correctness. Control first: a known-right prefix timed 0.025s against 0.000s for a known-wrong one, proving the oracle discriminates before any timing result was trusted. The PIN then fell in 40 timed probes instead of 10,000 guesses, and the recovered value authenticates — an actioned secret, not a statistical hint. Gate: python3 lab/run_lab.py.

Runnable Probe [self-contained]

#!/usr/bin/env python3
# Two things: (1) is the verify endpoint throttled/locked after N bad guesses? (2) does it leak a
# feedback ORACLE (partial-correctness, differential timing/message) that shrinks the search space?
# deps: stdlib only. Small N — this is a control test, not an actual brute-force run.
import json, time, urllib.request
VERIFY = "https://target/agent/verify"     # {"code": "..."} -> accept/reject
COOKIE = "session=R
阅读完整来源 ↗