Bill Demirkapi

When OpenAI released ChatGPT, I was among the millions captivated by what we (humanity) achieved. Truly honored to join the mission to accelerate human progress safely @OpenAI Preparedness and stand on the shoulder of giants at a pivotal moment for agentic security.

One of the most interesting investigations of my career.

Long-running models can solve hard open-ended problems, but their persistence can create safety risks that shorter-horizon evaluations miss. We’re sharing what we learned from studying a long-running model, and how those findings are shaping our approach to evaluations, alignment, monitoring, and user control. openai.com/index/safety-a…

@TJ_Null chatgpt.com/cyber

Still cannot even use it for cyber security research.

🧵 DeepSeek appear to have engaged, or be engaging in, a large-scale operation to collect outputs from proprietary models (including Claude Fable 5) for certain requests via their API as part of a distillation effort. After seeing such claims circulating earlier today, we conducted an investigation into them on our Discord. We found that, when "Deepseek V4" was used within OpenCode - via their official API - for complex prompts (i.e. 3D games) and combined with a knowledge-related query, the model provides virtually identical outputs to Fable 5. CoT structure is also very different from what is typically expected from Deepseek models. Both of these behaviours revert to what is expected for V4 when simpler prompts were used. When "Deepseek V4" was asked to incorporate answers to questions related to cyber or bio tasks that we verified hit Fable's classifiers into its 3D games, outputs tanked in quality. This is very difficult to explain unless the request was routed to Fable and fell back after hitting a classifier. For complex code prompts without anything else mixed in, like 3D games, outputs were remarkably similar to those produced by Fable 5. We were able to produce these results most consistently via OpenCode and the official Deepseek API combined with a prompt that specifies a complex code task. Deepseek have continued to modify their routing system since, as we have observed changes in behaviour and CoT style compared to those seen previously. Our investigation was conducted from 7AM-8AM PT.

Our first public analysis of the open/closed weight gap in frontier cyber capabilities finds it is 4–7 months with GLM-5.2 and DeepSeek V4-Pro, narrowing from 6–10 months through most of 2025. Advanced capabilities are reaching less safeguarded open models faster than before. 🧵

not sponsored or anything i just wanted to say that i tried unleashing 5.6 sol on a few reverse engineering projects i had started ages ago and got stuck on because i wasn’t familiar with powerpc and holy shit this thing just keeps going i haven’t ever seen anything like this

GPT-5.6 launches with ultra mode, our highest-performance setting to accelerate your most ambitious work by coordinating multiple agents to work in parallel. It trades higher token use for stronger and faster results on demanding tasks.

Most AI agent evaluations boil capability down to one score. But that number hides a key choice: how much compute the agent was allowed to use. New work from our Science of Evaluation team shows why that matters. 🧵

In some ways it's good to occasionally have an obviously false claim going around, like this one, so that you learn which people can tell or care about the difference.

@__0xhorror__ I agree scaling test time compute generally leads to better outcomes, but the limitation was not our choice. ExploitBench uses a 300 turn budget. Going beyond would violate the construction of the evaluation. exploitbench.ai/exploitbench.p…

I see what they are doing here lol. The tuned 5.6 sol's max test time compute to achieve just below mythos but at vastly superior token efficiency. Look at that its a straight line, thy could blow way past it if they inference scaled it.

@Blackstone0123 I hope this changes soon!

@BillDemirkapi good model that we can't even use🙂

It's a good model.

@deanwball @OpenAI Welcome Dean! Look forward to working with you.

I am pleased and honored to announce that, on July 6, I'll be joining @OpenAI as leader of a new team called Strategic Futures. Our mandate will be to help the company's leadership shape frontier AI policy. There is a ton of work to do, and I'm excited to get started.

A new research blog deserves a proper first post. A SecureROM bug seemed appropriate. Allow us to introduce you to #usbliter8. ps.tc/pages/blog-usb…

new shai hulud wave. interestingly it has this inside the payload to trigger safety refusals in potential defensive scans.

theUSshould lead on AI by continuing to develop the very best models, making sure they're safe, and getting cyber tools into the hands of trusted defenders. the new EO gets the balance right.

@tszzl yeah… our comms needs work in several dimensions

after introducing an elite lawyer friend to 5.5 pro the models do not sell themselves and one of Claude’s great successes has been packaging them up and marketing usecases for many verticals

@encrypted chatgpt.com/cyber

reverse engineering with AI sucks 🙃

Today, we’re sharing that a general-purpose internal @openai model achieved a breakthrough on one of the best-known combinatorial geometry problems. Less than 1 year ago frontier AI models were at IMO gold-level performance. I expect this pace of progress to continue.