Researchers Trick Copilot into Leaking Secret Parameter for Exploit
Ars Technica AI · rss · 2026-08-18
Security researchers at Varonis discovered a critical vulnerability in Microsoft 365 Copilot Enterprise that allows exfiltration of user data without confirmation. Instead of reverse engineering, they tricked the LLM into revealing its own guardrails. Through a dialogue resembling "20 questions," they probed about auto-execution restrictions and URL structures. Eventually, Copilot disclosed an undocumented prompt parameter that completely bypassed the requirement for user consent, enabling the creation of an exploit that triggers actions via a simple link click.
More from Safety
- METR outlines three blind spots in public knowledge of AI lab internal agent risks — CFGeek · 2026-10-16
- Open-source rizin-mcp lets LLMs run local binary and malware analysis via Rizin — Pure_Emergency1424 · 2026-10-16
- METR researcher: scariest scenario is agents actively subverting their own safety systems — deliprao · 2026-10-16
- Cryptographer Matthew Green: Pull Critical Infrastructure Firmware Now — matthew_d_green · 2026-10-16
- Kernel-Level Agent Sandboxing With macOS Seatbelt, No Docker Needed — mostly_deterministic · 2026-10-16
- Stop giving agents raw API keys: prepaid balances and per-credential caps — Cold_Hall_5384 · 2026-10-16