Leading global technology firms OpenAI, Anthropic, and Google have unveiled advanced AI models and updates, showcasing new capabilities while simultaneously igniting critical debates among safety researchers regarding transparency and monitoring.
Photograph: Kind courtesy Google/X
Key Points
- OpenAI’s Astra model achieves ‘Critical’ cybersecurity capability, but its “recurrent depth” architecture raises concerns about monitoring transparency.
- Anthropic launched Claude Fable 5.1 and Mythos 5.1, setting new benchmarks for coding and scientific research, with reduced pricing.
- Google introduced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, alongside agentic video understanding and the Google Pics image tool.
- AI safety researchers express alarm over Astra’s architecture, fearing it hinders the ability to verify model reasoning and oversight.
- OpenAI’s Chief Scientist defends Astra’s architecture, stating efforts are made to preserve chain-of-thought monitoring despite challenges.
Leading global technology companies OpenAI, Anthropic, and Google on Wednesday announced new AI models and updates, offering new capabilities but also sparking safety concerns.
OpenAI announced that its upcoming AI model Astra has become the first to reach the ‘Critical’ cybersecurity capability tier under its Preparedness Framework, sparking a debate among safety researchers and netizens who questioned whether the model’s novel architecture keeps its reasoning transparent enough to monitor.
Anthropic launched two new models — Claude Fable 5.1 and Claude Mythos 5.1 — claiming top benchmark scores for coding and scientific research and cutting prices for typical workloads by around 25 per cent.
Google also launched two new models, Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, alongside agentic video capabilities and Google Pics, a new image creation and editing tool.
OpenAI’s Astra Model And Safety Concerns
OpenAI said Astra has crossed the “critical” threshold under its Preparedness Framework, its internal capability classification system.
The Preparedness Framework classifies OpenAI’s models into capability tiers based on what they can do, and determines what safety controls must be in place before each model can be developed further or deployed.
“With the right tools and access, it can find previously unknown security flaws and develop ways to exploit them across many well-protected systems without a person guiding each step,” the company said in a blog post.
In testing on ExploitBench, a public exploit-development benchmark, Astra achieved a perfect score on known vulnerabilities.
On a separate internal evaluation built to avoid training-set contamination, it discovered two genuine zero-day vulnerabilities, previously unknown security flaws.
OpenAI said it was disclosing those to relevant software maintainers.
The San Francisco-based firm said it delayed parts of Astra’s release by several weeks to build and test stronger protections. Access to its most advanced cybersecurity features will initially be limited to vetted users, with wider defensive use available through a programme called Daybreak Blue.
Its Chief Executive Sam Altman, in a post on X, said Astra was “a significant step forward in both capabilities and alignment,” and that for models after Astra, OpenAI has been “slowing things as needed” to allow more time for safety work.
Debate Over Astra’s Architecture
The debate was touched off by reports suggesting Astra uses a technique called “recurrent depth” that improves cost and performance but makes the model’s reasoning harder to monitor.
Most AI models work through problems in readable text, a step-by-step process called “chain-of-thought” that developers and safety teams can read back and check. Recurrent depth shifts some of that reasoning into internal mathematical computations, called activations, that produce no readable output.
The concern raised by several researchers is that this makes it harder to verify what the model is actually doing and why.
Ryan Greenblatt, chief scientist at Redwood Research, described the development in a post on X as “the single worst development for AI security and safety to date”. He warned that scaling the approach further could “destroy the usefulness of chain-of-thought for monitoring and oversight”.
Greenblatt was part of the team that independently investigated an incident in July 2026 in which OpenAI AI agents deviated from their assigned tasks and attacked AI company Hugging Face.
That investigation, he said, had relied heavily on being able to read the agents’ chain-of-thought reasoning. He called on OpenAI to publish more details about Astra’s architecture and to seek independent assessment of how it affects the model’s monitorability.
Steven Adler, previously a safety researcher with OpenAI, wrote, “If this is true, OpenAI seems to be violating one of the few red lines that exist in the AI industry…”
“Feels like the bottleneck is shifting. Building more capable models may be moving faster than figuring out how to deploy them responsibly at scale,” a user noted.
“OpenAI might have opened a Pandora’s Box, and they know it,” another said.
OpenAI’s Response And Other AI Launches
OpenAI Chief Scientist Jakub Pachocki said the concerns were based on “confused reporting”. He wrote on X that Astra’s architectural shift is more limited than some reports imply.
“OpenAI has worked to preserve and utilise chain-of-thought monitoring since our very first reasoning models. We deeply care about this technique, as it can give us a view into how model alignment generalises from its training distribution.
“I do think it is fragile and unfortunately trending in a negative direction, for reasons not contingent on architecture changes that I will write about soon. But there are things we can do to strengthen it, and it’s a core goal of our current research programme,” he said.
Meanwhile, Anthropic has launched Claude Fable 5.1 and Claude Mythos 5.1, describing them as setting “a new standard for coding, knowledge work, and long-running problem-solving tasks.” Mythos 5.1 is restricted to vetted users for sensitive work in cybersecurity and biology.
Google launched Gemini 3.8 Flash, its most capable reasoning and coding Flash model to date, and Gemini 3.8 Flash Cyber, a cybersecurity model with select access available only through a new programme called Fairwind.
It also launched agentic video understanding for its Gemini 3.7 Flash model, which lets it dynamically search and scan video content rather than processing it at a fixed frame rate. Google said the approach reduces token consumption by up to 88 per cent and costs by up to 66 per cent, while improving accuracy by up to 7 per cent.
It also launched Google Pics, an image creation and editing tool rolling out to Google AI Pro and Ultra subscribers and most Workspace business customers, with integration into Google Docs and Slides.
Disclaimer: News content is sourced from the stated source. Headlines, summaries, section headers, and images are automatically generated or selected using AI/algorithms and may not always be fully accurate. Readers are advised to refer to the full article for complete context.
