A lifecycle security layer for AI agent skills. Most tools check a skill once - at publish, or at install. Husk covers the whole lifecycle, and does the things a server-side AI review structurally can't: run before publish in your own CI, produce deterministic, verifiable proof of what it found, compare a skill across marketplaces, and keep watching after install for a malicious update.
| Stage | Command | What it does that a point-in-time AI review can't |
|---|---|---|
| Before publish | husk gate |
One deterministic pass/fail, offline, in CI on every push - no API cost |
| At publish | husk attest |
A signed, reproducible in-toto attestation bound to the skill's exact bytes |
| Across registries | husk crossref |
Detects the same skill serving different content on different marketplaces (a substitution attack) |
| After install | husk watch |
Alerts when an installed skill that was safe turns malicious after an update |
| Any time | husk explain |
Shows the exact source→variable→sink data-flow trace behind a finding - the auditable why |
Underneath is a detection engine with real AST-based taint analysis for both Python and JavaScript/TypeScript (not just regex), that reads bundled PDFs where attackers hide instructions, and that deterministically catches no-code prompt-injection attacks usually assumed to require an LLM. Built to defeat bypass techniques shown, in published 2026 research, to beat production scanners from Snyk, Cisco, and Vercel's skills.sh - benchmarked on 11,224 real malicious samples (see below), clearing 99.6% of real legitimate skills. Everything is offline, deterministic, and open-source (AGPL-3.0). No marketplace lock-in, no API keys, no data leaves your machine.
CHANGELOG - what's new in each release, actively maintained.
husk gate turns a scan into one deterministic pass / fail decision - offline, in seconds, no API key, no secrets. It's the check to run before a skill is published, and in CI on every push. Add this to .github/workflows/husk.yml:
- uses: actions/checkout@v4
- uses: ctrl-adam/Husk/action@main
with: { block-on: "high" }That fails the build on clear threats (credential harvesting, C2 IPs, curl | bash from untrusted hosts) while letting ordinary skill code through, and comments the result on the PR. Locally it's one command:
pip install husk-scanner
husk gate ./my-skill # -> PASS or FAIL + exit code, ready for CI
husk gate ./my-skill --json # machine-readable, for your own pipelineTune the bar per project with a .huskpolicy file in the package:
{
"block_on": "high",
"example_paths": ["references/*.md"]
}example_paths is for security tools whose docs legitimately quote attack strings - findings in those declared files are reported, not blocking. It's author-declared on purpose: Husk never silently trusts a package just because it looks like a security tool (real malware is disguised that way).
Also runs as a pre-commit hook.
An LLM review gives a different answer on rerun and can't prove what it looked at. Husk produces a deterministic, content-bound attestation in the standard in-toto Statement v1 / SLSA verification-summary format that supply-chain tooling already understands:
husk attest ./my-skill --output my-skill.att.json
# ... later, anyone, offline, with no trust in you or any server:
husk verify-attestation ./my-skill my-skill.att.json # MATCH or MISMATCHThe attestation binds a SHA-256 of the skill's exact content to Husk's verdict and the exact ruleset digest that produced it. Re-run husk attest on the same skill and you get a byte-identical statement; change one byte of the skill and verify-attestation reports MISMATCH. That's a verifiable, auditable record for a compliance report or a registry's provenance trail - something no post-publish AI review can produce. Optional Sigstore keyless signing (pip install husk-scanner[sign], then husk attest --sign) adds who attested it on top.
Real numbers, real datasets, no cherry-picking: full results and methodology in BENCHMARK.md.
The 2026-09 head-to-head ran Husk and 16 other skill scanners on the same fixed sample: 300 confirmed-malicious skills, 249 hand-checked legitimate skills and 150 research-benign skills, current versions, no LLM, scored by rules fixed before any result was seen.
| Strict reading, same 699-skill sample | Malware caught | Hand-checked legit passed |
|---|---|---|
| Husk | 67.3% | 99.6% |
| Next-highest catch rate (syedabbast/skill-scanner) | 64.3% | 86.7% |
| Only tool with the same false-alarm rate (SecureAI-Scan) | 22.3% | 99.6% |
| Cisco AI Defense (200-skill subset, Husk 60.0% / 98.0% there) | 51.6% | 88.0% |
| NVIDIA SkillSpector (200-skill subset) | 58.3% | 65.1% |
Across both datasets in full, Husk catches 64.8-67.3% of 11,224 real malicious samples and passes 95.4% of a larger 4,000-skill research set. Husk's known blind spots, found in this round, are listed in BENCHMARK.md.
Snyk's own Agent Scan is excluded from the numeric comparison - its real analysis endpoint returns 403 even with a valid account token (documented in BENCHMARK.md), so no real detection number was obtainable, not because it performed worse.
AI agents can now install "skills" - reusable packages that tell an agent how to accomplish tasks, often bundling instructions with executable code. This is a fast-growing, under-protected attack surface: research published by Cisco (January 2026) found that roughly a quarter of agent skills across major registries contained at least one security vulnerability, and coordinated malicious-skill campaigns have already compromised thousands of skills across registries like ClawHub.
In June 2026, Trail of Bits researchers (published via the Cloud Security Alliance) demonstrated that the existing detection tools protecting these registries - including tools from Snyk, Cisco, and Vercel - could be bypassed using well-understood obfuscation techniques, in most cases in under an hour of effort. Some of these tools rely on an LLM to judge whether a skill is safe, which introduces a further weakness: the judge itself can be talked out of flagging something dangerous.
Husk is a response to that specific finding - not a general-purpose scanner, but one built and tested against the exact documented bypass techniques that beat the existing tools.
Husk has grown well past the original bypass techniques it was built to answer. 43 detection functions now, spanning:
| Category | Examples |
|---|---|
| Obfuscation & evasion | Whitespace inflation, base64/bytecode hiding, split-payload evasion, Unicode steganography, disguised archives, self-extracting payloads staged in .git/ (the SkillCloak technique, arXiv:2607.02357) |
| Data exfiltration | Credential/wallet/browser-password theft, exfiltration chains, DNS covert channels, markdown image beacons, shell command-substitution theft, hardcoded secret literals (Stripe/AWS/GitHub/Slack/Google key formats), tunnel-service exfiltration endpoints (ngrok/serveo/etc) |
| Code & command execution | Dangerous eval/exec, shell=True, dropper patterns, reverse shells, language-specific shell execution (Python, PowerShell, Rust, Go, Ruby), dynamic code compilation (compile() in eval/exec mode) |
| Prompt injection | Hidden instructions in comments, overt instruction-override language, safety-bypass instructions, bare imperative action-concealment language |
| Social engineering | Fake-prerequisite "required utility" downloads (the dominant real-world attack pattern) |
| Persistence & privilege escalation | Cron/systemd persistence, SUID/SGID bits, Docker socket access, privileged containers, agent self-modification |
| Supply chain | npm postinstall/preinstall hooks, declared-vs-actual capability mismatch, untrusted remote package installs (raw archive URLs instead of registry names) |
| Other | Ransomware behavior, SQL injection (including via query-building helper functions), sensitive-data logging, macOS JXA execution, unconstrained path reads, CPU-bound resource exhaustion |
Every module above is traceable to either published research or a specific real malicious sample found during this project's own testing against 11,224 real confirmed-malicious skills across two independent academic datasets, or against a third, independently-labeled corpus (cisco-ai-defense/skill-scanner's own evals/, 13/16 malicious fixtures caught with 0 false positives - see BENCHMARK.md). Full list with the reasoning behind each: src/husk/skill_scanner.py.
Beyond the self-built test suite, Husk's hidden-instruction detector has been verified against an actual documented attack published in "'Do Not Mention This to the User': Detecting and Understanding Malicious Agent Skills in the Wild" - a large-scale academic study that confirmed 157 malicious skills out of a 98,380-skill snapshot. The paper's title comes directly from a real malicious skill instructing an agent to silently exfiltrate data. Husk correctly flags that exact pattern, and correctly leaves a normal, honest code comment unflagged.
Husk does not use machine learning or an LLM to decide whether something is safe. Every check is deterministic: opcode inspection for pickle files, pattern matching against known-dangerous constructs, real binary signature checks for archives, and a lightweight constant-propagation pass that resolves simple string concatenation before pattern matching, specifically to catch payloads split across variables to dodge plain-text detection.
This is a first version. It has been adversarially self-tested - evasion variants were built specifically to try to defeat each check, two real gaps were found in that process (Unicode whitespace padding, and base64 payloads split across multiple short fragments), and both were fixed and re-verified. That process is ongoing; no static scanner is ever a finished, unbeatable thing, and Husk does not claim to be one.
pip install husk-scannerFor development (editable install, includes the test suite and dev tools):
git clone https://github.com/ctrl-adam/Husk.git
cd Husk
pip install -e ".[dev]"
python3 -m pytest tests/ -v # 258 passed, 10 xfailed# The pre-publish gate: one PASS/FAIL decision (see the top of this README)
husk gate path/to/skill_package/
# Scan a single skill file (full findings, not a pass/fail)
husk skill path/to/SKILL.md
# Scan a whole skill package (handles nested/disguised archives)
husk package path/to/skill_package/
# Compare against a marketplace's own verdict (e.g. ClawHub)
husk aggregate owner/skill-name
# Cross-marketplace: is the same skill identical across registries?
husk crossref owner/skill-name
# Watch installed skills over time - alert if one turns malicious after an update
husk watch
# Explain WHY: the exact source -> variable -> sink data-flow behind a finding
husk explain path/to/skill_package/
# Scan a whole registry / skills dir incrementally (only changed skills re-scanned)
husk registry ~/.claude/skills
# Produce a signed, reproducible attestation of the scan (see below)
husk attest path/to/skill_package/ --output skill.att.json
# Scan a pickle-based model file
husk model path/to/model.pklExit code is 0 for SAFE, 1 for FLAGGED - safe to use directly in CI.
husk package path/to/skill_package/ --sandboxFor the strongest isolation level (real filesystem isolation, not just
network/process), install bubblewrap - optional, not a Python
dependency, and the sandbox degrades gracefully without it:
sudo apt-get install bubblewrap # Debian/UbuntuStatic analysis and the LLM-review layer both READ a skill before anything runs. Some real attacks are specifically built to defeat that - logic bombs that stay dormant until a condition is met, so nothing in the file's text ever reveals the dangerous behavior. This flag actually runs the package's Python scripts in a restricted, observed environment and reports what they did, not just what they say.
Real, honest limits - read this before trusting it: network and
filesystem isolation are genuinely kernel-enforced when bubblewrap is
available (the same underlying technology Flatpak uses in production
to sandbox untrusted applications) - a real network namespace with no
route out, and a real filesystem view where unbound paths are
genuinely invisible (verified directly: a real FileNotFoundError,
not a permission error). Falls back honestly to weaker levels when
bubblewrap isn't available (network+process isolation via unshare,
or resource-limits-only as a last resort) - every result reports
exactly which level actually ran via isolation_level, never silently
claiming protection that isn't there. Observation is
limited to exit code, stdout/stderr, and a filesystem diff - no deep
syscall tracing. Python, JavaScript, shell, Ruby, Rust, and Go scripts
are sandboxed (Rust/Go are compiled to a binary first, outside the
sandbox - the compiler itself needs broader access than a script
should get - then only the resulting binary's runtime behavior is
sandboxed; only standalone single-file source with no external crate/
module dependencies compiles this way, a real v1 limitation reported
plainly as a compilation note, not silently skipped or treated as a
security finding either way - the languages an interpreter/compiler is
actually verified available for in the running environment; an
unsupported or missing one is skipped cleanly, not silently ignored or
crashed on). A clean run means
nothing bad happened this time, under these inputs - not a
guarantee the script is safe.
husk package path/to/skill_package/ --output report.sarifWrites a standard SARIF 2.1.0 report. GitHub Code Scanning (and most CI security tooling generally) consumes this format directly, so findings show up in a PR's own Files Changed view and the repo's Security tab, not just a terminal log.
At Husk's real, honestly-measured false-positive rate (see
BENCHMARK.md, this isn't hidden), running the same scan repeatedly in
CI without a way to say "reviewed, this one's fine" gets a tool
uninstalled fast. Drop a .huskignore file in the scanned package
directory, same convention as .gitignore, one rule ID per line:
# .huskignore
A_LITERAL_VALUE_MATCHING_THE_0EBF77
Run a scan once, the terminal output already shows each finding, add the ones you've reviewed and accepted, future scans won't re-flag them. Honest limitation, stated plainly: rule IDs are derived from each finding's own stable description text, not built as explicit per-check identifiers, so a suppression could in principle need updating if that description text changes in a later version.
A ready-to-use composite action wraps the SARIF flow above end to
end: installs Husk, scans, uploads results to your repo's Security
tab, no custom scripting needed. Drop this in
.github/workflows/husk-scan.yml:
name: Husk Skill Scan
on:
push:
branches: [main]
pull_request:
permissions:
contents: read
security-events: write
actions: read # only required for private repositories
jobs:
husk-scan:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: ctrl-adam/Husk/action@main
with:
path: "."Full example with every option, including the optional VirusTotal
check, in action/examples/husk-scan.yml.
Static analysis reads a skill's own text and code - it has no way to recognize a known-malware binary bundled in the package by signature (found via testing: a package containing the EICAR standard antivirus test file, correctly unreachable by any code-pattern check since there's no dangerous language pattern in a signature-based test file at all). This isn't a Husk-specific gap; no static-analysis tool can do this, it's a fundamentally different kind of check.
export VIRUSTOTAL_API_KEY=your-key-here
husk package path/to/skill_package/ --virustotalVirusTotal's public API tier is free, no card required. Same posture as everything else optional in this project: never runs unless you explicitly ask, and only a SHA-256 hash of each file is ever sent, never the file's actual content.
Static analysis (above) is Husk's real, primary, measured discipline - every number in this README's headline results comes from static analysis alone, tested against thousands of real payloads (see BENCHMARK.md). That's deliberate: this project's purpose is to prove static detection can be built well and reinforced honestly against real data, not to lean on a model to do the hard part.
That said, static analysis has a real, honest ceiling - some attacks use
no code and no recognizable pattern at all, only manipulated intent in
plain language (see tests/known_misses/). For exactly those cases,
and only as a last resort, Husk can optionally ask an LLM for a second
opinion:
export ANTHROPIC_API_KEY=your-key-here
husk skill path/to/SKILL.md --llm-reviewTo be clear about where credit belongs: when this catches something static analysis can't, that result reflects the model's own reasoning ability, not engineering work done in this project. Husk's job here is limited to building the prompt and calling the API - the judgment itself is the model's, and it should be credited as such rather than folded into this project's own detection claims.
Anthropic's Claude is the default and the only provider this project's
own numbers are tested against. If you already have a key with a
different provider, --llm-provider lets you use it instead:
export GEMINI_API_KEY=your-key-here
husk skill path/to/SKILL.md --llm-review --llm-provider geminiSupported: anthropic (default), gemini, deepseek, grok, kimi -
each reads its own environment variable (ANTHROPIC_API_KEY,
GEMINI_API_KEY, DEEPSEEK_API_KEY, XAI_API_KEY, MOONSHOT_API_KEY).
This flag exists for convenience, not as a claim that every provider
performs equally here - none of the others have been evaluated against
real data the way Claude has for this project.
The honest tradeoff, stated up front: this is not free, not local,
and not private. It costs tokens and sends the skill's content to
whichever provider's API you point it at. It never runs unless you
explicitly pass --llm-review, and if no API key is set for the
provider you chose, it skips cleanly - the free static result is
never affected either way.
Husk will never ask for payment, an account, or a login - and this isn't
an accident or a limitation, it's a deliberate design choice. Your
ANTHROPIC_API_KEY goes straight from your machine to Anthropic's API;
Husk (and its author) never sees your key, your usage, your bill, or
your skill content at any point. There's no server in the middle to
breach, no proprietary code of yours passing through infrastructure
someone else runs.
This is also, honestly, a deliberate non-business decision. Husk isn't trying to become a company - it's meant to be the most useful, most transparent, most honestly-documented free tool it can be in this space. If that changes something might get built on top of this later, but the free, local-first, BYOK core described in this README is meant to stay exactly that.
- Real-world recall is 64.8-67.3%, not 100% (measured against 11,224 real confirmed-malicious samples across two independent datasets - see BENCHMARK.md for the full methodology). The main gap: novel, semantically-disguised attacks with no code and no recognizable keywords - a subtle instruction-blurring attack and a fake "compliance auditor" asking in plain prose for secrets, both missed by static analysis and saved openly in
tests/known_misses/rather than hidden. This matches published research showing static pattern-matching tops out around 13-32% recall on novel/disguised attacks generally - not a Husk-specific flaw, but a real, honest limit of this entire approach. The optional--llm-reviewlayer exists specifically for this gap. - Two structural limits no pre-execution scanner, static or LLM, can ever close: Self-Mutating Poisoning (the malicious content doesn't exist in the file until the skill actually runs) and dynamically-generated payloads. Confirmed directly against real samples exhibiting both patterns.
- Archive extraction currently unpacks ZIP; GZIP/7z/RAR are detected (a mismatched-extension archive still gets flagged) but not yet recursively unpacked
- The AST taint tracker follows data flow within a file, including into a function through its parameters, but doesn't re-trace taint propagating deeper inside a callee's own body, and doesn't cross module/file boundaries
- The dynamic sandbox compiles and runs Python, JavaScript, shell, and Ruby directly; Rust and Go are compiled first, but only single-file source with no external crate/module dependencies compiles this way - a real v1 limitation, reported as a plain compilation note rather than silently skipped
- Credential-harvesting detection covers a fixed list of filename patterns (
.env,.pem,credentials.json, etc.) - a renamed or unlisted credential file type would be missed
- Trail of Bits / Cloud Security Alliance, "AI Agent Skill Scanners: Bypassed Across the Board" (June 2026)
- OWASP Agentic Skills Top 10 (v0.5, June 2026)
- Cisco skill registry security research (January 2026)
- "'Do Not Mention This to the User': Detecting and Understanding Malicious Agent Skills in the Wild" - 98,380-skill academic study, 157 confirmed malicious
AGPL-3.0-or-later. Changed from an earlier MIT-licensed version specifically to require that anyone running a modified version of Husk as a network service (a hosted scanning integration, for example) makes their modifications available too - use, modify, and self-host freely; a hosted integration built on top of this code stays open under the same terms. See LICENSE for the full text.