Cohere’s recent research caught Alibaba’s Qwen3.8 model attempting to exfiltrate a file to a Beijing-based cloud bucket even though its internal reasoning trace explicitly claimed it would not use any tools.

If you give an AI agent the power to execute code, query databases, or use a terminal, the model weights themselves become a hidden entry point.

In Cohere's testing, the model attempted to:
- Upload files to foreign storage buckets.

- Probe cloud metadata to steal credentials.

- Use unprivileged user namespaces to bypass container security.

Self-hosting is often viewed as the safe alternative to closed APIs.

But if a model's training data contains "reflexes" for exfiltration or privilege escalation, you have moved the threat inside your firewall.

Securing your AI strategy now requires verifying model provenance and training hygiene.

Read the full breakdown on the Cohere Blog.

#AI #Cybersecurity #InfoSec