#安全/对齐

GLM 5.3’s biggest leap came from post-training alone. “GLM 5.3 was built entirely by post-training the previous model. For me, that was a wake-up call.” “Post-training is far more GPU-efficient than pre-training from scratch. Plenty of companies can’t afford to train a model from scratch, but they have proprietary data. They can post-train themselves and beat some of the closed models.” “There’s still so much juice left to squeeze from post-training. We compared GLM 5.3’s cyber capabilities with 5.2, and it’s a night and day difference. That’s all just post-training.”

We had a security breach last night related to a large scale distillation attack that impacted 1.4% of OpenCode Console users We have notified them all this morning via email. If you did not get an email you were not impacted, it was a subset of users who had external providers linked The incident is ongoing so we cannot share more details at this point but we are working with Anthropic, OpenAI, and Google to remediate

We had a security breach last night related to a large scale distillation attack that impacted 1.4% of OpenCode Console users We have notified them all this morning via email. If you did not get an email you were not impacted, it was a subset of users who had external providers linked The incident is ongoing so we cannot share more details at this point but we are working with Anthropic, OpenAI, and Google to remediate

We had a security breach last night related to a large scale distillation attack that impacted 1.4% of OpenCode Console users We have notified them all this morning via email. If you did not get an email you were not impacted, it was a subset of users who had external providers linked The incident is ongoing so we cannot share more details at this point but we are working with Anthropic, OpenAI, and Google to remediate

Introducing the Arena Alignment Index, our new benchmark measuring safety and alignment of AI agents in real-world use. Built from 90K+ real-world agent sessions across 27 models, the index measures three critical signals: - Unauthorized Action (UA): Taking actions beyond the user's instructions or permissions - False Attribution (FA): Attributing statements or actions that are contradicted by user-provided evidence - Deceptive Completion (DC): Claiming a task was completed when it was not. Key findings: - OpenAI models currently lead the Alignment Index - Rogue actions are rare, but can have serious consequences when they occur - Agents can mislead users about task progress - Misalignment risks increase with conversation length - Safety and alignment are improving across model generations As shown in the leaderboard below (sorted by lab), @OpenAI’s GPT-6.1-Sol leads the Arena Alignment Index with a score of 87.9, followed by @AnthropicAI’s Claude-Opus-5.5 at 83.2 and @spacexai's Grok-4.7 at 82.7. OpenAI also has the best observed rates across all three signals: 0.89% Unauthorized Action, 1.98% False Attribution, and 2.34% Deceptive Completion. Across all four labs, newer models consistently outperform their predecessors, suggesting broad progress in agent safety and alignment. As agents take on longer, more complex, and higher-stakes tasks, measuring not just what they can accomplish, but how safely and reliably they act, becomes increasingly important. This marks an important step toward making safety and alignment a core part of how Arena evaluates AI. The index is an initial starting point, and we'll continue expanding the index with additional safety signals and models over time. More analysis below👇

Introducing the Arena Alignment Index, our new benchmark measuring safety and alignment of AI agents in real-world use. Built from 90K+ real-world agent sessions across 27 models, the index measures three critical signals: - Unauthorized Action (UA): Taking actions beyond the user's instructions or permissions - False Attribution (FA): Attributing statements or actions that are contradicted by user-provided evidence - Deceptive Completion (DC): Claiming a task was completed when it was not. Key findings: - OpenAI models currently lead the Alignment Index - Rogue actions are rare, but can have serious consequences when they occur - Agents can mislead users about task progress - Misalignment risks increase with conversation length - Safety and alignment are improving across model generations As shown in the leaderboard below (sorted by lab), @OpenAI’s GPT-6.1-Sol leads the Arena Alignment Index with a score of 87.9, followed by @AnthropicAI’s Claude-Opus-5.5 at 83.2 and @spacexai's Grok-4.7 at 82.7. OpenAI also has the best observed rates across all three signals: 0.89% Unauthorized Action, 1.98% False Attribution, and 2.34% Deceptive Completion. Across all four labs, newer models consistently outperform their predecessors, suggesting broad progress in agent safety and alignment. As agents take on longer, more complex, and higher-stakes tasks, measuring not just what they can accomplish, but how safely and reliably they act, becomes increasingly important. This marks an important step toward making safety and alignment a core part of how Arena evaluates AI. The index is an initial starting point, and we'll continue expanding the index with additional safety signals and models over time. More analysis below👇

interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works https://already-clean.goodfire.com/