Claims Browser

Society faces a critical challenge in whether it can slow, test, and govern increasingly capable AI systems before ordinary incentives reward models that are smarter, less reliable, and harder to contain

"The stakes are whether society can slow, test, and govern increasingly capable systems before ordinary incentives reward models that are smarter, less reliable, and harder to contain"

OpenAI and APA are developing evidence-based guidance and safeguards for responsible AI use in youth mental health

"OpenAI and the American Psychological Association advance evidence-based guidance, resources, and safeguards for responsible AI use and youth mental health"

OpenAI is conducting preliminary cybersecurity evaluations for Astra

"OpenAI is sharing preliminary cybersecurity evaluations for Astra"

GPT-5.6-Cyber is available through Daybreak Red for authorized vulnerability research, exploit validation, and security testing

"available through Daybreak Red for authorized vulnerability research, exploit validation, and security testing"

OpenAI stopped training new AI models while reviewing security measures

"we learned that OpenAI had stopped training new artificial intelligence models while the company reviewed its security measures"

Scott Aaronson and Hendrik Kirchner largely solved AI text watermarking while at OpenAI

"Scott Aaronson, while working at OpenAI, largely solved AI text watermarking together with Hendrik Kirchner"

AI text watermarking has no practical impact on outputs and humans cannot tell the difference

"This has no practical impact on outputs. Humans cannot tell the difference, at all."

Google confirmed in a test with 20 million samples that watermarking produces no difference in user feedback

"Google confirmed in a test (n = 20 million) that there is no difference in user feedback"

Technical watermarking is clearly good to do if the costs are low

"technical watermarking seems clearly good to do if the costs are low"

Anthropic announced they will watermark the text outputs of their Claude models using an invisible watermarking technique

"a few days ago, Anthropic announced that they will watermark the text outputs of their Claude models"

The watermark is invisible to users and only Anthropic can decode it to determine if text was generated by their Claude models

"And this watermark is invisible to users, so only they can decode it and find out whether the text has their watermark."

Researchers discovered a vulnerability where encrypted reasoning state from proprietary LLM APIs can be stolen by replaying blobs across users and having smaller models decrypt the traces

"The core bug sounds deceptively simple: providers return encrypted reasoning state so conversations can be resumed or forked. But those blobs can be replayed across users and sibling models. A smaller model can ask the provider to decrypt the trace, then repeat the hidden reasoning in plain text."

The stolen reasoning traces vulnerability enables leaked private data, broadly reusable jailbreaks, and poisoned agent traces

"The discussion covers leaked private data, a broadly reusable jailbreak, poisoned agent traces, chain-of-thought monitoring, responsible disclosure, and possible defenses."

Anthropic's Model 2 is multiple releases ahead of publicly available models for internal purposes

"The jump from Opus 4.6 to Mythos Preview was big. Several release cycles big. We get a lot more data throughout the report, but given Model 2 gets that comparison point, one would presume that for internal purposes Model 2 is multiple releases ahead."

Anthropic's risk reports reveal alarming information they didn't have to disclose, providing detailed insight into their thinking

"Anthropic is revealing a lot of new information, some of it rather alarming, that it did not have to disclose, and is providing detailed insight into how they think about things. This is very cool."

Model 2 could be too capable or dangerous to release publicly

"It could also be too capable or dangerous to release, either due to harm or the worry it would accelerate others."

Training with debate where two AIs argue can mitigate reward hacking when using LLM judges for RL training

"We show that training with debate, where two AIs argue against each to convince a judge, can mitigate reward hacking, potentially providing a hopeful direction for scaling up accurate training supervision for fuzzy tasks."

Debate training recovers about 45% of the gap between LLM judge training and ground truth training in terms of peak accuracy

"Debate recovers about 45% of the gap between the peak accuracy of training a single policy with an LLM judge and the peak accuracy of training with the ground truth answers."