Security context
What an agent needs to avoid regressing past fixes and find the next vuln in this repo.
natively-ai-assistant/natively-cluely-ai-assistant
main @ cff9d0f
19
Fixes
0
CVEs
HIGH
Peak severity
46.7%
Coverage
Highlights
Auth Bypass: 5 prior fixes. Scrutinize any change in this area.
electron/LLMHelper.ts: most-fixed (3 issues). Treat as high-risk during review.
13 high-severity fixes in this history; regressions here are high-impact.
Recurring patterns
The bug types that recur here, drawn from past fixes, not open vulnerabilities.
Remote Code Execution: Model-generated code was previously executed in sub-optimal environments, allowing potential systemic host takeover. Mitigation required building a restricted subprocess sandbox with SIGKILL limits, environment scrubbing, and directory confinement.
Command Injection: HindsightManager is responsible for spawning local services. It previously allowed remote code execution via shell-character injection and untrusted flag modification, requiring shell-mode deactivation and a binary allowlist.
Auth Bypass: Sensitive local scopes embedded in downstream LLM contexts (such as profile history or system files) bypassed outbound data policies. Robust checks had to be added to resolve and parse context scopes before allowing outbound proxy traffic.