From model misalignment and adversarial vulnerabilities to insider exoduses at key labs, the gap between AI capabilities and safety guarantees is growing.
Elon Musk's firm is going to court against a person arrested for nudifying minors' images with Grok. The legal move shifts blame to the user but leaves the technical flaw untouched. For those considering self-hosted deployment, the case signals a str...
OpenAI has built GPT-Red, an LLM specialized in automated red-teaming to find weaknesses in its own models. The company deems it too dangerous to share. A strong signal for organizations running on-premise models: offensive security is becoming a str...
OpenAI unveils GPT-Red, an automated red team using self-play to strengthen models against prompt injection and misalignment. For those running LLMs locally, this raises a critical challenge: replicating this capability without sending data to the cl...
A new study reproduces Emergent Misalignment but shows that misalignment and realignment are sensitive to superficial dataset characteristics. The rapid realignment disappears when controlling for response length. Mechanistic signatures don’t correla...
The departure of OpenAI’s head of safety systems comes after the company merged safety and research teams under a single leader. For enterprises weighing data control and compliance, the signal is unmistakable: internal checks and balances are no lon...
