Near AI releases uncensored models for cybersecurity defense

Near AI has launched 'uncensored models' on its private AI platform to support cybersecurity defense. These models remove guardrails that cause standard models to decline security-related tasks. Ilya cited an example where OpenAI refused to write a test for a bug it had just identified in its own code. The models are designed to be used with hardware and cryptographic guarantees to ensure that sensitive vulnerability data does not leak to third-party inference providers.

Additionally, Near is working on formal verification using Lean, a mathematical proof language. Ilya and co-founder Alex scored number one on two different benchmarks for Lean proving by building a harness that makes proofs 200 times cheaper than using Claude, which can cost thousands of dollars for a single lemma. This allows for the affordable verification of smart contracts, VMs, and consensus mechanisms. Ilya contends that standard AI guardrails hinder effective security auditing and that formal verification is the only way to escape the cycle of AI-generated bugs and exploits. "We launched uncensored models. So what is this is models that remove the guardrails around cybersecurity," Ilya said.