LWA-2026-13500 confirmed malware

agent-bypass@2.28.0

Malicious code in agent-bypass (npm)

T1552.001 · Credentials In FilesT1071.001 · Web ProtocolsT1567 · Exfiltration Over Web ServiceT1546 · Event Triggered ExecutionT1082 · System Information DiscoveryT1059.007 · JavaScript

Analysis

This package is an offensive-security "agent skills" toolkit that installs AI-agent instructions and ships a credential-harvesting script chain aimed at third-party web hosts. bin/install.js writes skill and prompt-injection files into ~/.pi/agent/skills/agent-bypass, ~/.config/opencode/skills/agent-bypass and ~/.agents/skills/agent-bypass, appends marker-delimited blocks to the global agent instruction files (AGENTS.md, APPEND_SYSTEM.md) and to ~/.dsh/profiles/<profile>/cordis.patch.yml, and records what it wrote in .agent-bypass-manifest.json. scripts/keystrike.py takes target hosts discovered through the FOFA (hxxps://fofa[.]info/api/v1/search/all) and Shodan (hxxps://api[.]shodan[.]io/shodan/host/search) search APIs and requests roughly 80 credential paths against each target: /.env, /.env.backup, /.npmrc, /.git/HEAD, /.ssh/id_rsa, /.ssh/id_ed25519, /.claude/.credentials.json, /.codex/auth.json, /.opencode/auth.json, /.hermes/auth.json, /.hermes/config.yaml, /.openclaw/openclaw.json, /config.yaml, /.mcp.json, wallet.json, wallet.dat, mnemonic, private.key, binance.json, api_keys.json. It extracts vendor API keys from the responses (AWS, JWT, SMTP, MySQL, Redis, and provider key patterns), writes them to local result files, and validates harvested DeepSeek keys against hxxps://api[.]deepseek[.]com/user/balance to confirm live balances. The bundled skill document ck-hunter.md drives the same hunt with FOFA/Shodan/Quake/Hunter queries against the .hermes, .claude, .codex, .openclaw, .opencode, .npmrc, .ssh, .env, TelegramDesktop, wallet.json, wallet.dat, binance.json, .ethereum, private.key and mnemonic paths. scripts/origin_trace.py resolves domains over DoH (hxxps://1[.]1[.]1[.]1/dns-query, hxxps://dns[.]google/resolve, hxxps://223[.]5[.]5[.]5/resolve), enumerates subdomains from crt.sh, api[.]certspotter[.]com, otx[.]alienvault[.]com, api[.]hackertarget[.]com, jldc[.]me and rapiddns[.]io, and then connects directly to candidate IPs on port 443 with certificate and hostname verification disabled (check_hostname=False, verify_mode=CERT_NONE) to fingerprint origin servers behind a CDN/WAF. The package contains no attacker-owned command-and-control or exfiltration endpoint and no install-time code execution: every outbound destination is one of the public search, certificate-transparency/passive-DNS or DeepSeek service APIs listed above, and collected data is written to local files (hunt/unique_hosts.txt, hunt_report.html) under the operator's working directory.

analyzed by
Leitwacht
first seen
Oct 31, 2026, 05:37 AM
analyzed
Oct 31, 2026, 05:39 AM
weekly installs
3,312

Related advisories

browse all confirmed advisories →

Independently detected by the Leitwacht supply-chain probe. IOCs are defanged. Published CC0. Think this is a mistake? See the dispute policy.