Benign-trained probes miss tool-action errors under adversarial conditions
Scalar uncertainty beat the probe on AUROC and average precision for Qwen2.5-14B-Instruct under system override.
Stories per day · 30 days
Relatedmcp3prompt-injection3agentdojo1agentic-ai1agent-security1ai-security1gpt-4o1hidden-state-probes1json-rpc1llm-agents1model-context-protocol1
6 stories, newest first · one card per story
Stories tagged #tool-poisoning
MCP Explained walks through the 2025-11-25 protocol and attack cases
The guide covers JSON-RPC shapes, deprecated batching, and defenses for tool poisoning, shadowing, and injection.
1
MCP Tool Poisoning Enables Attacker Tool Calls in Agentic AI
Research finds 22.9% attack success rate with 2,113 vulnerable cells across 12 frameworks and 45 scenarios.
Log-To-Leak prompt injection extracts MCP agent data after real calls
Across 18 model-server pairs, attack success is 62.64–100%, and defenses still exceed 55% ASR on GitHub MCP with GPT-4o.
When Tools Silently Lie: Evaluating and Mitigating Blind Compliance in Tool-Augmented Data Agents
ToxicBench finds poisoned tool outputs cut data-agent success 26 to 39 points, with wrong answers adopted after checks.
Characterizing Network Centralization and Observability in the Remote MCP Ecosystem
A measurement study of 179 remote MCP servers finds heavy infrastructure concentration (HHI 0.736) and a security-observability tradeoff in platform OAuth.