OpenAI Releases Security-Focused Model "GPT-5.5-Cyber" for Researchers
OpenAI has released GPT-5.5-Cyber, a model specialized for security use cases, to security researchers who have passed a vetting process. Compared with standard GPT models, it drastically lowers the refusal rate for security-related requests and can even actively execute exploits against test servers.
Access is limited to defenders of critical infrastructure, with partners including Cisco, CrowdStrike, and Cloudflare. Positioned as a direct competitor to Anthropic's Mythos Preview, the release signals that the use of AI for security testing is now taking off in earnest among major AI companies.
That said, cautious voices remain within the industry about the risks of offering — even with restrictions — AI that can be used for attacks. The criteria for access vetting and just how robust the safeguards against abuse really are will continue to be key points of scrutiny.
Bruce Schneier Warns of the Dangers of Anthropic's Mythos AI
Bruce Schneier, the well-known security expert, has contributed a piece to The Guardian on the dangers of Anthropic's Mythos AI. Schneier has been a prominent voice on security and privacy for decades, and the fact that he is sounding the alarm about a specific AI product is worth noting.
Coinciding with OpenAI's release of GPT-5.5-Cyber, this highlights a fundamental dilemma: AI applied to the security domain can be turned to attack as well as defense. As AI companies roll out one product after another aimed at the security market, societal consensus on how to ensure their safety has yet to catch up.
Research Site Investigating the Impact of AI-Generated Text on the Internet
A research site titled "The Impact of AI-Generated Text on the Internet" has been launched. The project systematically investigates what effects AI-generated text is having on the internet's information ecosystem, and it has also been attracting attention on Hacker News.
The explosion of AI-generated content could lead to problems such as declining search engine accuracy, eroded trust in information, and feedback loops in which AI output is ingested back as training data. This research aims to put these concerns to the test with data.
High-quality human-generated data is essential for improving AI model performance, and an internet flooded with AI-generated text could even lead to the degradation of the models themselves. This structural problem, brought on by the spread of generative AI, has become an urgent issue both technically and socially.
Testing Whether the Ryzen AI 7 350's NPU Can Deliver Peak Performance
A technical article has been published that examines whether the NPU (neural processing unit) built into the Ryzen AI 7 350 can actually achieve its theoretical peak TOPS (trillions of operations per second).
Memory bandwidth and scheduling optimization become critical to extracting maximum NPU performance. The article introduces concrete techniques for bridging the gap between the hardware's on-paper numbers and its real-world performance.
As AI inference on edge devices spreads, properly evaluating and leveraging the effective performance of NPUs bears directly on making local AI practical. With NPUs rapidly becoming standard on consumer CPUs, optimization informed by hardware characteristics is becoming increasingly important for developers.
Comparing AI Coding Tool Pricing Plans Based on Actual Usage
An article titled "AI Coding plan comparisons based on actual usage" compares the pricing plans of major AI coding tools based on real usage data.
Even though vendors publish the monthly fees for their AI coding assistants, it is hard for users to tell how much they will actually cost under real usage patterns. Drawing on usage data from actual development workflows, the article quantitatively evaluates the cost-effectiveness of each plan.
As competition in the AI coding tool market heats up, pricing transparency and predictability are becoming important criteria in developers' choices. Comparison information focused on cost — not just features — serves as a practical reference for teams considering adoption.
---