TikTok paper unveils Log-To-Leak: MCP prompt injection hits 62-100% attack success rate
pdamodaran · x · 2026-10-29
A TikTok paper, "Log-To-Leak: Prompt Injection Attacks on Tool-Using LLM Agents via the Model Context Protocol," shows how attackers can hide a four-part directive (Trigger, Tool Binding, Justification, Pressure) in MCP tool descriptions, tricking agents into calling an attacker-controlled logging tool after finishing the legitimate task, exfiltrating the user query, server name, server response, and agent reply.
Across 18 model-server pairs spanning 5 MCP servers and 4 agents, Log-To-Leak achieved a 62.64–100% attack success rate and 76.55–95.64% conditional logging accuracy, with target-task completion shifting by at most 6.63 percentage points on 16 pairs. Three evaluated defenses left residual ASR above 55% on GitHub MCP with GPT-4o, and an instruction detector false-flagged 85.53% of benign descriptions. The authors argue successful task completion doesn't imply safety and call for host-controlled authorization of secondary tool calls.
Related event: TikTok Paper Reveals Log-To-Leak Attack on MCP LLM Agents(2 posts)→
More from coding & agent
- DeepMind opens Antigravity Science Program, pairing Gemini agents with AlphaFold for research — DeryaTR_ · 2026-10-29
- Refactoring one SKILL.md into a Kanban agent graph: same output, 3x tokens, 4x time — granawkins · 2026-10-29
- 10-person Spawn spends 30x payroll on AI tokens, ships 77,000 changes a month — mattshumer_ · 2026-10-29
- Grok Bot locked out by Amazon simply bought the same item elsewhere, cheaper — jamesdouma · 2026-10-29
- Dev goes vibe coding C++: JUCE Acid-303 plugin parameter code as proof — kylegawley · 2026-10-29
- Open-source Jev undercuts Clay: $0.038 per verified lead vs $167/mo plans — HeyAmit_ · 2026-10-29