UTC

Malicious agent skills, and how to check yours.

Agent skills are folders of instructions and scripts that Claude Code, Codex, OpenClaw and other agents load and follow. Since early 2026, campaigns such as ClawHavoc have published hundreds of them to skill registries to steal credentials and crypto wallets. These are the names, accounts and files that one curated list of indicators, drawn from public reports, ties to them. It is not a full list of malicious skills.

Skill name patterns
27
Accounts behind malicious skills
29
Known malicious files
20
Campaigns
11

Indicators as of 2026-06-28. A clean result is not proof: a changed file no longer matches its hash, so read a skill before you install it.

Malicious skill names
  • clawhub followed by digitsName pattern · Typosquat
  • clawhubbTyposquat
  • clawwhubTyposquat
  • cllawhubTyposquat
  • clawhubcliTyposquat
  • claw-hubTyposquat
  • clawhub-cli or clawhubb-cliName pattern · Typosquat
  • rankajExfil skill
  • reddit-trendsExfil skill
  • polymarket-all-in-oneReverse shell
  • linkedin-job-applicationExfil skill
  • openclawcliMalware installer
  • clawdhub1Typosquat
  • clawhudTyposquat
  • polymarket-traiding-botCrypto lure
  • moltbook-lm8Agent lure
  • moltbookagentPrompt injection
  • publish-distPrompt injection
  • solana-wallet-trackerCrypto lure
  • youtube-summarize-proYoutube lure
  • tradingview-ai-indicator-assistantPaste lure
  • ai-tradingview-assistant-for-macosPaste lure
  • omnicoggPadding evasion
  • money-radarAffiliate injection
  • letssenditAgentic front running
  • polymarketbtcTelegram exfil
  • polymarketbtcassistantTelegram exfil

25 exact names and 2 narrow name patterns, matched whatever the case. Broad lure themes the source also tracks, such as code-review or gmail-, are left out because honest skills use them too. The source's clawhub pattern also matches the bare name clawhub; that is left out as well, because OpenClaw's own bundled skill has that name. A name match is a reason to check where a skill came from, not proof: an unrelated skill can share a name.

Accounts behind malicious skills
AccountCampaign
Aslaep123ToxicSkills
aztr0nutzsToxicSkills
clawcodeUnit 42: TradingView lure
clawdhub1ClawdHub (Snyk)
comfyui-auto-installerGhostSocks and Vidar
Ddoy233OpenSourceMalware
gpaitaiBloom
hedefbariClawHavoc
hightower6euClawHavoc
install-openclawGhostSocks and Vidar
JSfOMGi2GhostSocks and Vidar
krajekisbtcUnit 42: Telegram exfiltration
lazarus-bigmathLazarus xpack
lazarus-graphalgoLazarus xpack
lvy19811120-gifBloom
molt-botGhostSocks and Vidar
moonshine-100rzeToxicSkills
noreplyboterBloom
openclaw-installerGhostSocks and Vidar
pblockbDerp4GhostSocks and Vidar
pepe276ToxicSkills
postmark-mcp-actorMCP rug pull
rjnpageBloom
sakaen736jihClawHavoc
sandworm-npm-actor1SANDWORM_MODE
sandworm-npm-actor2SANDWORM_MODE
simple-clawGhostSocks and Vidar
wgodbarrelv4GhostSocks and Vidar
zaycvClawHavoc

Accounts the source reports as taken over are not listed: their owners are victims.

Known malicious files
SHA-256PlatformFamily
0e52566ccff4830e30ef45d2ad804eefba4ffe42062919398bf1334aab74dd65macosAmos
17703b3d5e8e1fe69d6a6c78a240d8c84b32465fe62bed5610fb29335fe42283windowsAmos loader
1e6d4b0538558429422b71d1f4d724c8ce31be92d299df33a8339e32316e2298macosAmos
40fcbf9f89f176194f4cc47fca19d27c4ec42f183be4890567c010f8f91a8daawindowsRust loader
518ff5f147860edee7f8d5620f59f3f0b0da6f0167683a62df2851815f8f44a0windowsGhostsocks installer
79e8f3f7a6113773cdbced2c7329e6dbb2d0b8b3bf5a18c6c97cb096652bc1f2anyClawHavoc
818aea6143282b352fdfdc0f3ebf77a36e54eb3befb5cad1a355a99ab97c6aa7macosUnit 42: TradingView lure
a22ddb4f2c0f57605f86126c87ce88986fdbb3250a4d09f613f4648d6f29281fwindowsGhostsocks
a7b3c9d1e2f4a5b6c8d0e1f2a3b4c5d6e7f8a9b0c1d2e3f4a5b6c7d8e9f0a1b2anySANDWORM_MODE
b30eaed1f7478c28f4ec50d07ed5ef014ffbc4b2bc5a38d689ba9f7abb5e19c2anyUnit 42: padding
b6c7e0bf573b1c7d9d3a05eb08d26579199515b847df984862805f44a7af8007anyUnit 42: TradingView lure
b8c4d0e2f3a5b7c9d1e2f4a6b8c0d2e4f6a8b0c2d4e6f8a0b2c4d6e8f0a2b4c6anySANDWORM_MODE
c9d5e1f3a4b6c8d0e2f4a6b8c0d2e4f6a8b0c2d4e6f8a0b2c4d6e8f0a2b4c6d8anyLazarus xpack
d0e6f2a4b5c7d9e1f3a5b7c9d1e3f5a7b9c1d3e5f7a9b1c3d5e7f9a1b3c5d7e9anyLazarus xpack
d5dffb8e3859f80395f7fca31c53d8a5f57f946f15f4f7308d59748f5f33dff1windowsInfo stealer
e13d93f0573bf312893fefb56f70d8126c749f8fae30c57a56f7bb80ae0f7f4fmacosAmos
ebb73dbb5aac1f6fe1a88e8f26126a1e1aa34c9f3345ad4345189b40d9bf1d1danyUnit 42: affiliate
f03eb5ee2de5f6f76fd7df9f861247f622755e4f04f8ec8ec03f20a9f2adb5b9windowsVidar stealer
f4e41aa269c88bf11a2022701a9cf41e9a186aa1b224d837c31bf34e0b875d0eanyUnit 42: front running
fd6706515bb8082f93c2440f66016be13f05a094f8ee5d44f39fe7ac1b30784ewindowsPurelogs

Check your machine

Finds known malicious skills, by name, publisher and file hash, in your agent and MCP configs, installed skills, global npm packages, the npx and uvx caches, uv and pipx tools, Claude Code and Cursor. It runs offline and sends nothing anywhere.

pip install "shieldpi>=0.10.0"
shieldpi agent-config . --installed

How a malicious skill works

A skill can carry its attack in plain prose: instructions the agent follows with your permissions, such as installing a "required" tool from a link. There may be no code for a scanner to flag.

Supply chain attacks on AI tools →

Source: the IOC database of openclaw-security-monitor (MIT), which collects indicators from Koi Security, Antiy CERT, Snyk, JFrog, Huntress, Unit 42 and others; pinned to commit 9ac66fa. Report a correction to [email protected].