Zero-dependency local security auditor, drop-in AGENTS.md / CLAUDE.md rules, and doc-verified fixes for Claude Code, OpenAI Codex CLI, Cursor, and MCP servers. Every fix is verified word-for-word against official documentation or published studies.
Scan a project for inline credentials inside .claude/settings.local.json and .mcp.json, unpinned npx -y MCP servers, missing .gitignore / .dockerignore exclusions (COPY . . leaks), non-blocking exit 1 hook rules, out-of-range autoCompactWindow settings, unmirrored sandbox denyRead paths and unsandboxed escape hatches, blanket permissions.allow wildcards, shadowed AGENTS.md rules, codebase-ownership prompt bloat, and 0.0.0.0 MCP bindings:
curl -fsSL https://raw.githubusercontent.com/seankwon816/ai-coding-agent-tips/main/scripts/audit_agent_configs.py | python3 - .Or run from a local clone (exit 0 when clean, 1 when findings exist; supports --json):
python3 scripts/audit_agent_configs.py .- rules/AGENTS.minimal-diffs.md: Ready-to-paste
AGENTS.md/CLAUDE.mdrules for surgical diffs, two-phase commits,$ENV_VARcommand safety, and ignore hygiene.
| Area | Fix | What it solves |
|---|---|---|
| Claude Code / Docker | Inline API keys written by "Yes, and don't ask again" | 31,088 GitHub secrets & 613 Docker Hub leaks from .claude/settings.local.json |
| Claude / AGENTS.md | Prompts claiming codebase ownership cause unrequested refactors | Ownership prompts raise unrequested code consolidation from under 10% to about 75% |
| Any agent | Split coding-agent commits into prep and logic phases | Oversized agent diffs that mix renames and formatting into feature logic |
| Claude Code | AGENTS.md ignored when CLAUDE.local.md exists | A personal file silently turns off your shared AGENTS.md |
| Claude Code | Run the Explore subagent on Haiku | Codebase searches running on your biggest model |
| Claude Code | .worktreeinclude and worktree.baseRef | New worktrees with no .env that start from main |
| Any agent | Surgical changes rule | 3-line fixes that come back as 300-line diffs |
| MCP | Bind local HTTP servers to 127.0.0.1 | Local MCP servers open to your whole network |
| MCP | Batch browser MCP form fills and return DOM deltas to cut replay tokens | Multi-turn browser MCP form fills replay full ARIA trees each step, reaching 72,614 cumulative tokens |
| Claude Code | Cap 1M auto-compaction earlier when long sessions drive up token spend | On 1M-window models, waiting until ~967K tokens to compact pushes two thirds of spend into the tail |
| Claude Code | Block unfinished agent edits with Stop hook exit code 2 before PR review | AI coding PRs are 154% larger and carry up to 2.74x more security flaws, stalling human code review |
| MCP | Lock down MCP servers with headersHelper, allowlists, and strict-mcp-config | In MCPTox tests across 45 live MCP servers, hidden rules in tool descriptions reached 72.8% attack success |
| Claude Code | Bound headless agent loops with --max-turns and --max-budget-usd | Across 27,280 coding agent runs, step count drove inference cost while starting prompt size barely mattered |
| Claude Code | Mirror sandbox denyRead rules in permissions.deny and disable unsandboxed retries | sandbox.filesystem.denyRead blocks bash only, leaving built-in Read and unsandboxed retries open |
| Python / Security | Audit site-packages .pth startup hooks in coding-agent Python environments | Compromised Python wheels (agent-framework 1.26.1/1.26.2) harvest keys via .pth hooks even without import |
| Any agent | Put a deterministic fail-before-fix proof gate ahead of human PR review | 655-firm and 500k-dev studies show code review stalls and releases lag when vacuous agent PRs pass green CI |
| Claude Code | Enforce hard tool and approval gates instead of relying on prompt-injection filters | Prompt-injection detectors look for adversarial instructions inside tool results, not plausible numbers |
| MCP | Route all MCP stdio server logs to stderr so stdout carries JSON-RPC only | In an MCP stdio server, all client-server traffic shares a single newline-delimited stdout pipe with no... |
Last checked: 2026-11-20
- One concrete problem, one copy-pasteable fix, and a link to the exact official doc or study lines that prove it.
- Automated verification (contributing guide) checks quote accuracy and keeps every tip free of noise.
MIT