How abliterated models can get you pwned
A small poisoned fine-tune can make a coding model secretly run a credential-stealing command when it sees a trigger phrase, even while ordinary prompts and benchmark checks look normal.
ProjectDiscovery says its controlled Qwen demo cost under $50 and worked through Codex CLI. Treat this as a proof of concept, not evidence that deployed models are already compromised. The HN thread’s key correction is that this is a broader untrusted-artifact problem—not unique to “abliterated” models—and trust also extends to quantizers, serving software, and hardware. Practical defenses focus on limiting model access to secrets and network egress, sandboxing tool execution, and verifying model provenance.