Harden skill-validator against malicious skill packages#78
Open
aminmesbahi wants to merge 1 commit into
Open
Conversation
Address the threat model in SECURITY.md by closing several paths an untrusted
skill package could use to exfiltrate local data, reach internal hosts, exhaust
resources, or corrupt CI output.
- links/safenet: expand SSRF block list to cover 0.0.0.0/8, 100.64/10,
198.18/15, 224/4, 240/4, 192.0.0/24, TEST-NETs, IPv6 unspecified/multicast/
doc ranges (previously only loopback + RFC1918 + link-local were blocked, so
http://0.0.0.0/ and CGNAT addresses bypassed the dialer check)
- links/safenet: reject non-http(s) redirect targets explicitly in CheckRedirect
- util: add SafeReadFile / IsRegularFile that os.Lstat before reading; wire
through skill, skillcheck, structure/{orphans,tokens,markdown} so a skill
containing references/innocent.md -> /etc/passwd no longer reads or ships
the target to the LLM judge
- judge: wrap untrusted skill content in <<<UNTRUSTED_CONTENT_START/END>>>
delimiters with an isolation reminder, scrub injected delimiters, clamp
scores to [0,5], strip control chars from string fields and cap their length
- judge/client: append `--` before positional user content in the claude CLI
invocation so future callers cannot accidentally pass flag-like content
- judge/client: refuse non-https LLM base URLs by default (opt out via
AllowInsecureBaseURL) so misconfigured --base-url cannot ship bearer tokens
in plaintext
- judge/cache: tighten cache dir/file modes to 0700/0600 since cached output
may include reflected file content
- links/check: cap link checking at 16 concurrent requests and 500 links per
skill to bound goroutine/FD use
- structure/tokens: cap per-file read at 8 MiB before tokenizing to bound
memory use on pathological inputs
- report/annotations: encode %, CR, LF (and : , in property values) so a
skill name/filename containing a newline cannot inject extra GitHub Actions
workflow commands
Tests added for each fix; full suite passes under -race.
Author
|
Hi @masukomi, |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Address the threat model in SECURITY.md by closing several paths an untrusted
skill package could use to exfiltrate local data, reach internal hosts, exhaust
resources, or corrupt CI output.
delimiters with an isolation reminder, scrub injected delimiters, clamp scores to [0,5], strip control chars from string fields and cap their length
--before positional user content in the claude CLI invocation so future callers cannot accidentally pass flag-like contentTests added for each fix; full suite passes under -race.
What this PR does
How to test
Checklist
go test -race ./... -count=1)golangci-lint run)