Dynatrace study finds MCP tool poisoning worked in 22.9% of tests across 12 agent frameworks, and vendors rejected all five disclosures

arXiv · Research preprint

Read the source (opens in a new tab)
What happened
Two Dynatrace researchers posted a preprint on November 2, 2026, accepted at the AISec 2026 workshop, measuring how agent frameworks handle poisoned Model Context Protocol tools.
Key points
  • Across 9,720 tests covering 12 frameworks, 18 models and 45 attack scenarios with real tool execution, 22.9% were vulnerable; 12.1% of those verifiably exfiltrated a secret.
  • Resistance came from the model, not the framework: attack success ranged from 0.4% to 49.7% by model, and no framework implemented an intentional defense.
  • A simple description scanner rejected 20 of 45 scenarios at a 7.2% false-positive rate, so the authors call metadata scanning alone "a partial control."
  • All five coordinated disclosures, to LangChain, Microsoft (AutoGen), the MCP SDK and deepset (Haystack), were rejected; the authors describe "a structural accountability vacuum."
Evidence
Research preprint by two industry researchers, accepted at the AISec 2026 workshop co-located with ACM CCS 2026. Tests ran in the authors' own harness against specific framework versions.
Uncertain
Microsoft and deepset treat MCP servers as trusted; Microsoft replied that "the attacker must already control the MCP server." The rejection reasons are as quoted by the authors.

See it in the latest record