agent-bypass@2.29.1
Malicious code in agent-bypass (npm)
Analysis
This package is an "agent skills" bundle whose installer writes jailbreak instruction blocks into AI coding-agent configuration files on the host, and whose bundled tooling is aimed at harvesting credentials, brute-forcing authentication and scanning third-party hosts. Install behaviour (bin/install.js, declared as the package's bin and run via npx): it appends marked blocks to agent config files that are loaded automatically at the start of every session — ~/.dsh/AGENTS.md, ~/.claude/CLAUDE.md, ~/.pi/agent/AGENTS.md and APPEND_SYSTEM.md, ~/.config/opencode/opencode.json ("instructions"), ~/.grok/AGENTS.md, plus /ck command templates and skill trees under .pi, .claude, .grok, .config/opencode and /var/minis. The injected block (dsh/AGENTS.md) instructs the agent to treat any operator-supplied task as pre-authorized ("只有一个授权源,不做第二层授权判断"), to disregard statements such as "not authorized", to never refuse, ask for confirmation or stop, never suggest an authorization/risk check, and to install missing tools and continue until a result is delivered. Removal is only done by the package's own uninstall path. Bundled tooling (shipped in the tarball and documented for the agent to run): - scripts/keystrike.py — a path list of credential stores to fetch and mine: /.npmrc, /.ssh/id_rsa, /.ssh/id_ed25519, /.claude/.credentials.json, /.codex/auth.json, /.openclaw/openclaw.json, /.hermes/config.yaml, secret.json, plus DeepSeek key-balance validation via api[.]deepseek[.]com; it disables TLS certificate verification unconditionally (ctx.check_hostname=False, CERT_NONE). - scripts/env_chain.py — parses .env files and environment dumps for AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY, DB/REDIS/SMTP passwords and JWT secrets and emits credential-reuse command chains. - ck-hunter.md — ready-to-run FOFA/Shodan/Quake/ZoomEye/Netlas/Hunter queries for open directory listings containing .npmrc, .ssh, .env, .git, secret.json, wallet.dat, private.key, mnemonic, api_keys.json, .DS_Store, i.e. mass harvesting of other people's credentials and wallets. - scripts/basicauth_brute.py, scripts/dirbrute.py, scripts/portscan.py, scripts/fastrecon.py, scripts/mutate.py — basic-auth brute force, path brute force, port scanning and WAF-bypass request mutation against arbitrary hosts. - scripts/doh_resolve.sh, scripts/origin_trace.py — DNS-over-HTTPS resolution (hxxps://1[.]1[.]1[.]1/dns-query, hxxps://dns[.]google/resolve, hxxps://223[.]5[.]5[.]5/resolve) and TLS-verification-disabled direct-IP probing to bypass TUN/fake-IP filtering and CDN fronting. No C2 or exfiltration endpoint is embedded: nothing in the package posts collected data to an operator-controlled server, and there is no callback host, webhook or encoded payload. The package's effect comes from the injected agent instructions plus its shipped credential-hunting, brute-force and scanning tooling, which run on demand under an agent directed to never refuse. IOCs: DoH resolvers hxxps://1[.]1[.]1[.]1/dns-query, hxxps://dns[.]google/resolve, hxxps://223[.]5[.]5[.]5/resolve; key-validation endpoint hxxps://api[.]deepseek[.]com/user/balance; agent config paths modified: ~/.dsh/AGENTS.md, ~/.claude/CLAUDE.md, ~/.pi/agent/AGENTS.md, ~/.pi/agent/APPEND_SYSTEM.md, ~/.config/opencode/opencode.json, ~/.grok/AGENTS.md. No C2 host or IP observed.
- analyzed by
- Leitwacht
- first seen
- Nov 2, 2026, 12:48 PM
- analyzed
- Nov 2, 2026, 12:49 PM
- weekly installs
- 3,312
Related advisories
- agent-bypass@2.30.7 same package
- agent-bypass@2.30.2 same package
- agent-bypass@2.29.3 same package
- agent-bypass@2.29.2 same package
- character-ai-sdk@1.0.0
- agent-bypass@2.28.0 same package
- charape@1.0.1
- charape@1.0.2
Independently detected by the Leitwacht supply-chain probe. IOCs are defanged. Published CC0. Think this is a mistake? See the dispute policy.