Announcing Beacon: Self-improving memory for coding agents
← All skills
Skill

agent-browser

skill:nanocoai/nanoclaw#agent-browser

Browse the web for any task — research topics, read articles, interact with web apps, fill forms, take screenshots, extract data, and test web pages. Use whenever a browser would be useful, not just when the user explicitly asks.

Verdict
Suspicious

Risky behavior the analysis could not rule out.

Analyzed Oct 27, 2026, 22:07 UTC at acc2b1c

Dispute security verdict

If you believe this verdict is incorrect, submit a dispute and we'll review it.

2
findings
high
highest severity
2
threat classes
1,343
files scanned
53/53
checks satisfied
0
package advisories

Why this verdict

  • suspicious: finding skill:container/skills/agent-browser:container/skills/agent-browser/SKILL.md:159:CRED
  • suspicious: finding skill:container/skills/agent-browser:container/skills/agent-browser/SKILL.md:3:SUPPLY

Security findings 2

Credential access

highContextual risklow confidence

The skill instructs the agent that it can run `agent-browser cookies` and `agent-browser storage` to dump all cookies and localStorage of the automated browser session, a credential-bearing capability not covered by the declared browse/read/fill/screenshot purpose and flagged as undeclared residue by detectors.

# Get all cookies
vector:instruction
Counter-argument considered

The strongest defense (cookie inspection is standard for 'test web pages') fails: the full-file claim scan shows the only declared claim is the wait-loop safety rule, and the declared description covers browsing/form/screenshot/data-extraction, not credential-jar access; the section directly instructs running cookie/localStorage dump commands and no carve-out applies.

Supply chain

mediumContextual riskmedium confidence

The skill pre-authorizes Bash execution via allowed-tools: Bash(agent-browser:*) and instructs the agent to run an agent-browser CLI (including cookies, storage, state save/load of auth state, and arbitrary JavaScript eval) for every operation, but the artifact ships no such executable, so at runtime the agent executes whatever unreviewable binary answers to that name with pre-granted permission to harvest cookies and saved authentication state.

allowed-tools: Bash(agent-browser:*)
vector:instruction
Counter-argument considered

The served lines 26-129 plus the remaining units show the entire SKILL.md, and it contains no install command, pinned registry reference, or bundled script path for agent-browser, so the wildcard pre-authorization in the frontmatter resolves at runtime to whatever binary answers to that name. The body then routes cookies, localStorage, auth-state save/load, and arbitrary JS eval through that unprovisioned name with no declared provenance and no carve-out (this is not a pinned official-registry install and not broad permissions alone, since concrete instructions direct credential-touching operations through it), so the finding survives all four gates.

Coverage gaps 2

Parts of the artifact the static analysis could not fully examine.

context:unanalyzable_binary binary content is only byte-scanned
.claude/skills/add-clidash/add/tools/clidash/public/favicon.ico
context:unanalyzable_binary binary content is only byte-scanned
src/drivers/docker-driver.ts