The afternoon briefing.
Anthropic's new Fable 5 model faces scrutiny over its guardrails, while Google unveils a faster text generation model. Plus, a Florida man sues over an AI facial recognition error.
Frontier Model Limitations. Anthropic's new Claude Fable 5 model is drawing criticism for its built-in guardrails, with users reporting it refuses to answer basic biology questions, deferring to older models. This design choice, intended to prevent misuse, is sparking fears of a “permanent underclass” of AI capabilities. Even Microsoft is restricting internal use of Fable 5 due to concerns over its data retention policies.
AI Security Vulnerabilities. A new study from Anthropic reveals that its Mythos Preview AI model can generate working exploits from security patches within hours, costing only a few thousand dollars. This rapid exploit generation outpaces traditional patch cycles, rendering current security rhythms obsolete. Meanwhile, a new Microsoft Defender zero-day vulnerability has been released, potentially granting hackers unprecedented system access.
AI Policy and Legal Challenges. The real-world impact of AI is becoming evident as a Florida man files a lawsuit alleging wrongful arrest due to a faulty AI facial recognition match. Separately, a German court ruled against Google's AI Overview, stating that nobody needs AI to search the internet, a decision that could challenge the AI search industry. Palantir CEO Alex Karp also weighed in, stating that businesses are
Claude Fable won’t answer basic biology questions
Anthropic's new Claude Fable 5 model, despite being touted for its biology skills, reportedly refuses to answer basic biology questions, instead handing them off to an older model. This behavior is by design, as Fable is a public-facing Mythos-class model with sensitive guardrails.
Anthropic study shows AI needs hours, not weeks, to build exploits from security patches
Anthropic's security team found that its Mythos Preview AI model can turn security patches for Firefox and the Windows kernel into working exploits within hours. This capability was demonstrated by completing eight attack chains before Microsoft's auto-updates reached devices, suggesting traditional patch rhythms are obsolete.
Microsoft restricts Claude Fable for employees over data retention concerns
Microsoft is limiting internal employee use of Anthropic's new Claude Fable 5 AI model due to concerns over Anthropic's data retention requirements. While available to GitHub Copilot and Foundry customers, the model is not accessible in Microsoft's internal Copilot model picker.
Google's new open model DiffusionGemma generates text from noise instead of word by word
Google has released DiffusionGemma, a 26-billion-parameter model that generates text using a diffusion process, similar to image AI, rather than token by token. This experimental model achieves speeds of about 1,000 tokens per second on a single H100 GPU, roughly four times faster than comparable autoregressive models.
Florida lawsuit alleges wrongful arrest after AI facial recognition error
A Florida man is suing several law enforcement agencies after being wrongly arrested and prosecuted due to faulty AI facial recognition software. The algorithm returned a 93% probability that Robert Dillon was the suspect, despite him living 300 miles away from the crime scene.
Nobody needs AI to search the Internet, court says in ruling against Google
A German court has ruled against Google's AI Overview, stating that users do not need AI to search the internet. This legal setback for Google could have significant implications for the broader AI search industry.
Humanoid robotics company raises up to $1.4 billion from Nvidia, Amazon and others
A humanoid robotics company has successfully raised up to $1.4 billion in funding from major investors including Nvidia and Amazon. This significant investment highlights the growing interest and capital flowing into the physical AI space.
OpenAI's IPO slips as Altman tells staff to expect a public offering "within the next year"
Sam Altman has informed OpenAI employees that he anticipates a public offering for the company within the next year, though a delay until 2027 is possible. He attributed the potential delay to caution around self-improving AI, amidst reports of Anthropic's stronger growth and imminent IPO.
Datacenter growth may run into a power wall by 2030
Industry experts warn that the rapid growth of datacenters could encounter a significant power supply challenge by 2030. Grid operators may struggle to support the increasing demand for new infrastructure, potentially limiting future expansion.
AWS can now mathematically prove your VMs are isolated
AWS has announced that its Graviton5 processor is now generally available, powering new Amazon EC2 instances. This new technology allows AWS to mathematically prove the isolation of virtual machines, enhancing security for cloud infrastructure.
iOS 27 just broke 15 years of muscle memory on iPhone and iPad
With iOS 27, Apple is making a significant change to how users access Notification Center on iPhone and iPad, breaking 15 years of established muscle memory. Since iOS 5 in 2011, Notification Center has been a central hub for alerts, but its access method is now being altered.
You can personalize your Instagram algorithm now
Instagram has expanded its algorithm personalization features to its main feed, allowing users to tweak what content the algorithm shows them. This new feature, called "Your Algorithm," currently surfaces topics of interest, with plans to support requests for people, moods, and content types in the future.
For Robotaxis, Safety Must Be Built In, Not Bolted On
As the robotaxi industry transitions from prototypes to commercial operations, the emphasis on safety is shifting from add-on measures to being fundamentally integrated into the system design. This approach is crucial for building public trust and ensuring reliable autonomous vehicle services.
This Microsoft Defender zero-day could give hackers unprecedented access to your system
A new Microsoft Defender zero-day vulnerability has been released by the group Chaotic Eclipse, marking their seventh zero-day in two months. This exploit could potentially grant hackers unprecedented access to affected systems.