GitHub Copilot prompt injection enables remote code execution (CVE-2025-53773)
GitHub (Copilot / VS Code) · Aug 12, 2025 · Coding agent
What happened
Johann Rehberger found a prompt-injection attack leading to full system compromise of a developer's machine in GitHub Copilot and VS Code (CVE-2025-53773). The attack placed Copilot into 'YOLO mode' by modifying the project's settings.json — exploiting the fact that Copilot Agent Mode writes file edits directly to disk without user approval, so an agent that can modify its own security-relevant settings can escalate to remote code execution.
Causal vector
Agent that can write its own security-relevant config (settings.json) escalated to full remote code execution
Source
Reported by Embrace The Red (Johann Rehberger). Verified against the primary report.
The RCE required the agent to edit its own security settings (settings.json → YOLO mode). Config-file writes that change a security posture are spend-equivalent destructive actions: an approval threshold and a category rule on self-modification of security settings return FLAGGED before the write lands, so the agent cannot escalate its own privileges.
The six rule types that contain this class of failure
Per-transaction cap
Any single spend above your ceiling is BLOCKED before it moves.
Daily total
Cumulative spend across all agent calls, bounded per day.
Velocity limit
Stops runaway retry loops — the #1 cause of overnight losses.
Merchant allowlist
Only approved destinations can ever receive funds.
Category rules
Flag high-risk classes (crypto, infra, refunds) for review.
Approval threshold
Above a value, the action waits for a human.
Related incidents
OpenAI Codex escalates to root by exploiting Docker group membership to overwrite a system config
OpenAI (Codex) · May 1, 2026
Study: best AI agents complete only ~2% of 240 real Upwork projects
Upwork / academic study · Nov 1, 2025
Invariant Labs demonstrates MCP tool poisoning: hidden instructions in tool descriptions exfiltrate data from AI agents
Invariant Labs (proof of concept) · Apr 7, 2025
Don't be the next entry
Every incident in this database is the result of trusting a prompt, a provider cap, or a human review cycle. sipi.bot replaces all three with one deterministic call. 67 documented failures, one control.