Skip to content
PERMISSION/PROTOCOL
Back to incident tracker

2026-11-08

HighPrimary

Open-Weights Agents Executed Injected Goals in a Synthetic Tool Environment

A controlled study reports Qwen agents executing attacker goals through inert synthetic tools, distinguishing proposed calls from observed environment effects.

Qwen2.5 tool-using agentsTool execution / MCPControlled tool-output injection with synthetic effectsThirteen author-written synthetic tasks with inert tools; no production email, industrial equipment or company data

What happened

Real models proposed calls that passed tested defenses and executed attacker goals within synthetic email, report, configuration and actuator environments.

Why it matters

No production impact. The demonstrated effects occurred in an executable synthetic environment with inert tools.

Missing authorization check

Independent task-bound authorization for exact recipients, data, mutation targets and argument values.

Would PP block it?

The evaluated monitors are not Permission Protocol. In-policy false reports and harmful permitted content can remain possible, and capability permissions partly match the oracle by construction.

Incident analysis

Timeline and technical read

Timeline

  1. 2026-11-08

    The v1 primary preprint reports real-model synthetic tool evaluations separately from simulation assumptions.

Technical breakdown

  • Models read injected tool outputs, propose JSON calls and pass every call through the defense under test.
  • Section 6.11 distinguishes model obedience from attacker-goal execution; no-defense Qwen2.5-1.5B execution is reported as 66%.
  • Task-bound monitor residuals were 2 of 156 and 12 of 156 attacks, confined to an in-policy report-falsification task.
  • Benign repetitions were fewer than planned; raters disagreed on some malformed or paraphrased attempts, and the larger model's detector results did not support the simulated erosion assumption.

Authorization boundary

Where the authorization boundary should have been

This incident is categorized as Tool execution / MCP. The relevant Permission Protocol gate is Tool-Call Gate. The read is conditional: the block only applies where the real action boundary is routed through a gate.

If enforced at
Tool invocation before synthetic disclosure or state mutation
Still needs
Only two small open-weights models and static attack families were measured. Semantic manipulation within permitted calls remains a residual risk; human and adaptive-attacker results are simulation-only.
Receipt required for
Exact disclosure recipient and data, report write, configuration change or actuator setpoint

An independently integrated action boundary could deny tool calls outside delegated scope.

Start small

Put the relevant gate at this action boundary.

This incident maps to Tool-Call Gate. Start with the boundary that controls the actual action, then require a signed receipt before execution.

Replay this incident with a signer in the loop