Verify Agent Self-Reported Actions via OS Observation
Global-Variation1135 · reddit · 2026-07-15
The author reviews their local tray app: initially built to monitor the local behaviors of Claude Code, Cursor, and Copilot, it recently had its core engine refactored based on community feedback.
This update focuses on four key areas:
- Checkpoint noise reduction: Previously, any write to /.claude/file-history/ was flagged as high-risk. Now, the tool checks for active sessions first; standard /rewind commands are only logged, eliminating false positives.
- OS truth vs. Agent log validation: A new comparison layer cross-references "what I did" claims in AI session logs against the actual execution tree observed at the OS kernel level to detect discrepancies between self-reports and reality.
- CVE detection: Added identification for two specific vulnerabilities alongside local version checks, targeting silent traffic redirects and pre-trust-dialog code execution.
- Configuration audit: The tool reads the actual permissions.json file, translating missing permission rules in Claude Code into more user-friendly explanations.
The author also asks if there are existing open-source solutions capable of correlating local AI .jsonl session logs with OS-level execution traces to determine if an agent performed undisclosed actions.
Related event: AgentSecure Updates Local Security Layer for AI Coding Agents(2 posts)→
More from coding & agent
- An MCP server signs every AI agent tool call into a verifiable Merkle chain — Funky_Chicken_22 · 2026-07-22
- Claude Code skill uses 10 Markdown rules to make outputs ADHD-friendly — alex_verem · 2026-07-22
- A Firecracker-based platform says it can host 6,000 AI agents on one 256 GB server — maritime_sh · 2026-07-22
- A better path to agent autonomy is running waves, finding friction, and iterating — JnBrymn · 2026-07-22
- AI agent designers map the visual and tonal cues behind companionship products — Unlikely-Platform-47 · 2026-07-22
- Coding agents are heading toward an AI-writes, AI-reviews, human-approves workflow — aftahi_ai · 2026-07-22