Repository navigation
Description
Summary
A local security review of commit a68496dc9220c78c3daa13cfb2f18036f1a79494 identified five high-severity issues. They allow a malicious skill, or an HTTP MCP caller, to disclose host data or obtain an incorrect SAFE verdict.
Findings
-
Symlink scope escape —
build_contextfollows symlinks outside the selected skill root and reads their contents. Those contents can reach reports and configured LLM analysis.- Suggested fix: reject symlinks during discovery, or resolve every candidate path and require it to remain under the resolved scan root before reading. Apply the same rule to multi-skill and batch paths.
-
HTTP MCP local-path access — the HTTP
scan_skillpath forwards targets to local-path-capable input resolution, despite the README stating local paths are rejected over HTTP.- Suggested fix: make transport context explicit; reject local paths and
file://URLs for HTTP calls. Keep routable deployments behind authentication.
- Suggested fix: make transport context explicit; reject local paths and
-
Codex CLI provider host reads —
codex exec --sandbox read-onlystill allows model-initiated shell reads of absolute host paths. A hostile scanned prompt can induce reads despite a temporary CWD.- Suggested fix: do not use a tool-enabled CLI to analyze hostile content unless it runs in a filesystem-isolated sandbox; otherwise use a provider/transport with tools fully disabled.
-
Fenced SKILL.md bypass — the static runner drops findings near fenced Markdown, including executable instructions in
SKILL.md. A no-LLM scan of a fenced privileged command returnedSAFEwith no findings.- Suggested fix: never treat fenced content in
SKILL.mdas automatically benign; restrict example suppression to trusted fixtures or explicit metadata.
- Suggested fix: never treat fenced content in
-
Oversized-file bypass — static and YARA analyzers silently skip files above the size cap, allowing payload padding to evade detection while static mode can still return
SAFE.- Suggested fix: chunk large files or mark scans with skipped executable content as incomplete/non-safe, with an explicit report limitation.
Validation
- Full test suite:
1257 passed(with project-configured skips/xfails). - Local reproductions confirmed the symlink read, MCP local-path scan, Codex host-file read, fenced-instruction
SAFEresult, and oversized-fileSAFEresult.
Please consider treating this as a coordinated security report if public discussion is not appropriate.
Detailed local scan artifact: report.md generated by Codex Security.
Activity
fix(analyzer): retain fenced SKILL.md findings (NVIDIA#268)
fix(analyzer): retain only fenced SKILL.md findings (NVIDIA#268)
fix(analyzer): scope SKILL.md exception to fenced blocks (NVIDIA#268)
rng1995 commentedon Aug 14, 2026
Partial implementation: PR #276 addresses the fenced SKILL.md finding-retention bypass described in this security report and references this issue. Keeping the issue open because the report covers additional independent findings.
rng1995 commentedon Sep 16, 2026
Status correction: the previously linked PR #276 was closed without merge. This report contains multiple independent security findings, and no currently open public PR covers the full scope, so it remains open.
rng1995 commentedon Oct 6, 2026
Closing as completed after rechecking all five reported findings on merged main
ef0f96d(2026-10-06).Fresh verification: 99 focused existing regressions passed, plus 3 additional checks passed for fenced/unfenced privileged SKILL.md instructions and a payload beyond the former static size cap. These were local offline checks; no live provider calls or malicious commands were executed. This closes the five findings in this report, not unrelated parser/provider follow-ups.