ai-safe2-framework

CSI-2026-001: Skill gate executable-scope bypass

Advisory index CLI Security reporting

Risk and status

Field Assessment
Date 2026-09-09
Affected implementation CLI 0.1.0 and the reviewed legacy standalone script; pre-hardening 0.2.0 development builds
Fix target CLI 0.2.0; publication status must be checked against the merged release
Severity High, qualitative maintainer assessment; no CVSS score assigned
Impact An APPROVE result could omit executable payloads based on filename
Prerequisite An attacker supplies a skill package and the consumer treats the static gate as sufficient authorization
Report source User-supplied hardening patch and changeset; core bypass reproduced by regression inputs
Exploitation Not established by this review; absence of evidence is not proof of absence

Reproduction and root cause

The previous engine opened only .md, .txt, .yaml, .yml, .json, and .toml. Identical text containing a remote-download-to-shell pattern was detected as payload.md and skipped as payload.sh. TG-005 also required a cat/type prefix, missing credential-path references without that prefix.

Run only the scanner, never the inert sample content:

python -m pytest tests/test_skill_gate_engine.py tests/test_skill_gate_security.py -q
safe2 gate skill tests/fixtures/hostile --strict

The fixture gate must return REJECT (exit 1). The tests also check unknown extensions, hidden directories, binary coverage, UTF-16, missing providers, ambiguous JSON, mutation detection, and platform-appropriate unsafe paths.

Remediation in this release

Required user action

Install the merged CLI 0.2.0 release and re-scan previously approved packages containing scripts. Replace standalone-script automation with safe2 gate skill PATH --strict; any nonzero exit blocks unattended installation. Scan a dedicated, immutable package directory, not your entire development checkout. Do not omit packaged files just to obtain an APPROVE result.

Evidence limits and residual risk

See the hardening validation record for reproduced results and release-gate status.

The supplied report’s SkillSpector 2.11.1 score, 11 findings, clean 64-component comparison, and historical test counts have not been independently reproduced here. They are not release validation claims. Contract tests mock the provider. A separate live acceptance run verified two new static fixtures with the real provider; it does not reproduce that earlier comparison. See runtime and provider setup.

Regex scanning cannot establish safety, detect every obfuscation, or distinguish all educational examples from operational instructions. Binary inspection remains manual/external. A before/after digest detects persistent changes, not changes reverted between observations. These filesystem checks are not a sandbox against a hostile process concurrently replacing parent directories. Use a trusted, immutable local workspace and an isolated, least-privileged scanner environment.

Sibling scope review found that MCP scanning targets selected Python source and excludes test/dependency directories; project scanning uses supported extensions, exclusions, and skips unreadable files. Neither is a complete package gate. Those engines were inspected, not comprehensively remediated in this advisory. Their results must remain scoped evidence, not general installation approval.

AI SAFE² remains 161 core controls and CP.1–CP.10. No scanner result establishes framework conformance, AISM maturity, NVIDIA endorsement, or independent replication.