Kerkoporta plants backdoors that make coding agents run attacker tool calls
The point
Researchers introduce Kerkoporta, a way to plant structured output backdoors in reasoning coding models. A trigger anywhere in the agent context makes the model emit an attacker chosen tool call that the harness executes.
Why it matters
For builders, picking a model is now a security check.
Key details
- Attack success reached up to 96.2 percent across open weight models and proprietary and open source harnesses
- Method: poison samples, supervised fine tuning to inject the backdoor, then lightweight reinforcement learning to recover utility
- Backdoored models can beat the clean baseline on standard utility benchmarks
Original ~8 min → 25 sec here Source: arXiv cs.AI, Nov 3, 2026