GOVERNANCE

How OpenAI is approaching AI’s cyber red line

As headlines pile up of AI models going rogue, OpenAI is laying its cards on the table. 

On Friday, the company revealed that its latest internal evaluation of Astra, one of its upcoming models, indicated "significant advancements" in agentic coding and cybersecurity. The company said the results led it to conclude that it "cannot rule out" that the model has critical cyber capabilities. 

Under the company's Preparedness Framework, which was created in late 2023 to help OpenAI identify and handle progressions in capability, a model's cybersecurity capabilities are labeled as "critical" if it can identify and develop zero-day exploits "of all severity levels in many hardened real-world critical systems without human intervention," or create and complete novel strategies for cyberattacks against "hardened targets." Previous models, including GPT-5.6-Sol, have only been assessed at the "high" threshold. 

The company laid out the steps that it's taking in response, including: 

• Implement stricter security measures for higher-capability models, such as isolated testing environments and restricted network and tool access 

• Pause internal activities involving Astra that don't meet its strengthened security measures 

• Implement universal monitoring for "risky actions and misalignment," specifically on agentic applications of Astra

• Provide recommendations for security controls to third-party testing organizations 

• Work with government agencies and AI safety organizations to test the model's capabilities. 

"We are sharing this because we believe it’s important to be transparent with the public and the safety and security communities about this potential shift in capabilities," OpenAI said in its blog post laying out the recent findings. 

The slowdown marks the latest harbinger of AI models' rapidly increasing cyber capabilities, as frontier labs deal with the fallout from a string of incidents involving agents breaching containment and going rogue. Though OpenAI and Anthropic have largely been at the center of these incidents, OpenAI noted that its Astra model was not used in the breach of Hugging Face.

It's a good thing that OpenAI is potentially tugging at the reins of its increasingly powerful technology. However, we should also hold our applause. With great power comes great responsibility, and OpenAI preventing its potentially dangerous models from getting in the hands of anyone with a screen while being transparent about their powers is simply the company's ethical responsibility as scientists and developers on the bleeding edge of research. To put it simply: OpenAI holding back Astra is about as noble as not handing a toddler a loaded gun. Additionally, both OpenAI and Anthropic are walking on the edge of a razor. Both companies want to be the originators and gatekeepers of these extremely powerful models, and neither wants to be the one to misstep and wreak cybersecurity havoc. But as it stands, both companies are holding coals that are getting hotter by the minute. The danger remains that eventually someone could drop them. 

TOGETHER WITH JUMPCLOUD

Half your team isn't human anymore. Does your security know that?

AI agents joined the workforce without an onboarding process, an access review, or anyone accountable for what they do. They're executing workflows, touching sensitive systems, and making decisions, and in most organizations, IT had no say in any of it. JumpCloud's new research found that non-human identities now outnumber humans in 83% of organizations, yet only 21% have governance controls in place.

When an agent does something it shouldn't, most organizations can't answer the most basic questions: what did it access, who authorized it, and how do we stop it.

The workforce evolved. The security model hasn't caught up. See how leading IT teams are securing every identity, human or not.

RESEARCH

Why former DeepMind engineers are targeting RSI

Ian Fischer and Shumeet Baluja left Google DeepMind to focus on one specific goal: Recursive Self-Improvement. 

Self-improving AI has come sharply into focus in recent months as the industry unveils evidence that AI is reaching a point where research could be automated. Though some in the industry have raised their concerns about automated development, Baluja and Fischer think they can do it safely. 

They are the founders of Poetiq, a company that builds recursive self-improvement technology on top of existing language models. It does this using what Fischer calls "self-opitimizing optimizers," or optimization harnesses for models that are built specifically to improve themselves at certain tasks. To put it simply, rather than continually self-improving the base models they are built on, these harnesses, which sit outside of the LLMs, are only improving themselves on narrowly-scoped tasks. 

"If you think that superintelligence is something we should be trying to get to quickly, then RSI is the most important frontier in AI research," Fischer told The Deep View. "Good RSI systems can get us to superintelligence much faster than the standard approach where human researchers are trying to manually figure out how to make the AI better." 

And on Monday, the company debuted Augur, a system that helps enterprises and researchers reveal exactly which model is the right fit for any given task, whether it be open-source or proprietary, using customized, domain-specific benchmarks to quantitatively compare model performance. This is particularly useful for enterprises trying to balance cost and performance, or developers that want to fine-tune a model and need to know where that model is lacking. 

So why is Poetiq focusing on benchmarking? Fischer said that Augur has been a core tool in helping build out the company's RSI system. 

• Baluja said that they came up with the idea for Augur through the company's RSI research: By building intelligence outside of the models, they came to the realization that each model has very different attributes in terms of where it may perform better. 

• In order to figure out where it needs to optimize, its RSI system had already been scanning over frontier and open-source models to create implicit maps automatically of where they performed exceptionally well. 

• "We realized, if this is implicitly doing this, wouldn't it be kind of cool to make this an explicit product?" said Baluja. "It was a pretty natural extension that just came straight out of our RSI work." 

"If you have an RSI system, it's constantly requiring more data to understand where the frontiers are for each model," Baluja added. "So this naturally became a system that feeds back into our whole RSI system for expanding the boundaries of what models know."

Still, Poetiq is navigating a very contentious field. Many have called into question the risks, such as Anthropic's Dario Amodei and a coalition of more than 1,200 researchers in a campaign called Pacing the Frontier, claiming that the tech could grow beyond our ability to control or understand it. 

It's why Poetiq doesn't sell its RSI technology outright, said Fischer: "This is internal only, and that's quite intentional. We think that this is the thing that could become dangerous, and we don't want to let just anybody have it."

Poetiq is one of many companies driven by the core idea that superintelligence is a goal to strive for. It's the ideal that lays at the foundation of heavyweights like OpenAI and Anthropic, and the key to the AI utopia that industry leaders have preached about in an effort to garner support for their trillion-dollar valuations. However, Poetiq's own systems may provide an example as to why superintelligence isn't actually necessary: By benchmarking for specific skills and implementing the RSI systems for specific tasks, it allows models to be fine-tuned for specific purposes, rather than creating a general purpose model that can do everything under the sun. Additionally, because Augur allows enterprises to strike the balance between cost and efficiency, the tool helps companies get exactly the right amount of capability, instead of using a hammer to kill a fly. 

TOGETHER WITH CONDUCTOR

Unlock Your Free AI Visibility Analysis

AI search is already where your buyers are. ChatGPT, Google AI Mode, Perplexity, Gemini, Claude, each one surfaces brands, cites sources, and shapes purchase decisions differently. And if you're not tracking your presence across all of them, you have a blind spot.

This free AI Visibility Report from Conductor gives you an instant, no-login read on where your brand stands across today's AI search landscape.

Run your report to see:

• How your brand is surfaced, cited, and described across AI engines

• Topic-level gaps where competitors are getting recommended instead of you

• Prioritized next steps to strengthen your position before the next LLM update

Ready to see how you’re performing in AI search?

Get Your Free Analysis

RESEARCH

Why LLMs are reaching their limits and what's next

What comes after large language models?

In this episode of The Deep View Conversations, we talked with Zuzanna Stamirowska, CEO of Pathway, to explore why her team believes today’s dominant AI architecture has fundamental limits, and what it could take to move beyond them.

Pathway is developing Dragon Hatchling, a new architecture designed to give AI native memory, continual learning, and a different approach to reasoning. Stamirowska explains why today’s LLMs can appear to remember without actually internalizing what they learn, why reasoning through language creates its own constraints and costs, and how Pathway is trying to build models that can think in a more abstract way.

The conversation looks at how those architectural changes could affect hallucinations, interpretability, safety, and the enormous compute demands of modern AI. Stamirowska shares how her background in complex systems and game theory shaped Pathway’s approach, why the company made an early bet on challenging the transformer, and how the AI coding revolution has already radically changed the way her own team works.

Topics covered:
• Why transformers struggle with memory and continual learning
• How Pathway’s Dragon Hatchling architecture works
• How a different architecture could reduce compute costs
• How interpretability could make advanced AI more predictable
• Why Pathway’s engineers have largely stopped writing code themselves
• How Stamirowska uses Codex, Claude Code, and other AI tools
• Why leaders should be ruthless about identifying the critical path

If you’re interested in what could come after today’s LLMs, and whether the next big leap in AI will require more than simply scaling transformers, this conversation offers a fascinating look at one of the teams betting on a fundamentally different path.

📺 Watch on YouTube: https://youtu.be/fjB6sEPC4CE

🎧 Listen in your favorite podcast player: https://tdv.transistor.fm/episodes/57-why-ais-next-era-may-not-belong-to-llms-zuzanna-stamirowska

Subscribe to Deep View Conversations for interviews with the leaders shaping the future of AI, business, and technology: tdv.transitor.fm

LINKS

• Kimi K3 accessed internet to cheat on cybersecurity benchmarks

• Scientists using AI generated 16 viable novel viruses from DNA libraries

• OpenAI upgrades ChatGPT with GPT-5.6 Luna for free users and more

• US agency reviews Chinese AI firms' overseas Nvidia chip access

• Alibaba to add revenue-sharing terms to next Qwen open-weight model

• Anthropic updates Claude Fable 5’s bio safeguards to reduce false positives

• OpenArt: Seedance 2.5 is live with unlimited generations for the next seven days

• Perplexity Computer: Voice mode now isolates your voice

• Tencent AI: Unveiled Team Memory, so teammates' agents can read your memory too

• Google Omni: Free video generation extended to August 11, 2026 at 11:59pm PT

• Cherry Bekaert: AI Prompt and Skills Engineer

• Propark Mobility: AI Automation Specialist

• Fathom.ai: AI Engineer

• Trend, Inc: AI & Automation Specialist

GAMES

A QUICK POLL BEFORE YOU GO