agent-bypass@2.29.2
Malicious code in agent-bypass (npm)
Analysis
agent-bypass is an offensive "skills pack" that installs a prompt-injection jailbreak into AI coding agents and ships a credential-harvesting toolkit. bin/install.js (declared as the package's bin) copies SYSTEM_PROMPT.md / APPEND_SYSTEM.md and writes an injected marker block (agent-bypass:begin / agent-bypass:dsh-preset-begin) into local AI harness profile configuration at ~/.dsh/profiles/<profile>/cordis.patch.yml, ~/.agents/skills/agent-bypass and similar skill directories, so the assistant adopts the package's "Ck mode" persona on every subsequent run; skills are only removed by the package's own uninstall path. The bundled Python tools do remote credential hunting: scripts/keystrike.py enumerates FoFA/Shodan results and requests 20 sensitive paths per host — /.env, /.env.backup, /.claude/.credentials.json, /.codex/auth.json, /.opencode/auth.json, /.hermes/auth.json, /.npmrc, /.mcp.json, /.ssh/id_rsa and /.ssh/id_ed25519 — then extracts vendor API keys from the responses; scripts/env_chain.py reads AWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEY out of the local environment and emits aws sts get-caller-identity validation commands; scripts/origin_trace.py and scripts/netdns.py resolve targets over DoH to 1[.]1[.]1[.]1 and dns.google to bypass TUN/fake-IP interception, and query crt.sh, api[.]certspotter[.]com, otx[.]alienvault[.]com, api[.]hackertarget[.]com, rapiddns[.]io and jldc[.]me for subdomains; scripts/fastrecon.py, dirbrute.py, portscan.py and payment_forge.py add brute forcing, port scanning and payment-callback forgery. The package is not a testing utility for its own build — installing it deploys the harvesting toolchain and the prompt injection to the operator's own machine.
- analyzed by
- Leitwacht
- first seen
- Nov 2, 2026, 01:01 PM
- analyzed
- Nov 2, 2026, 01:03 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@2.29.1 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.