The frontier model itself becoming a weapon — that report sent ripples through the security community. Three researchers reportedly used Claude Opus 5 to demonstrate a breach of an OpenAI-connected service in under 72 hours and for a few thousand dollars in token costs. On the product side, OpenAI launched a law-specific Astra while Google overhauled the foundation of its managed agents. This issue rounds up these moves along with funding news and topics from the local LLM scene.
Researchers Demonstrate OpenAI Internal Repository Breach Using Claude Opus 5 — 72 Hours, a Few Thousand Dollars
The biggest security news this week was a series of posts that spread on X (formerly Twitter) citing a Wall Street Journal report. Three researchers reportedly used Claude Opus 5 to chain together an image upload bug, the takeover of ChatGPT/Codex accounts, and access to connected services, finally proving access by sending a pull request to an OpenAI internal monorepo. The effort reportedly took under 72 hours and cost only a few thousand dollars in tokens.
What deserves attention is not the abstract argument that 'AI-powered cyberattacks are scary,' but the fact that what was exploited were 'ordinary integration surfaces' — SSO, forums, email, and business tools. An agent chaining known bugs and account takeovers step by step is a capability demonstration that is already materializing as a real-world threat. On X, concrete debates over control architecture unfolded, including privilege separation for instructions written by the AI itself and distinguishing sandbox failures from side channels. The thrust of the discussion seems to be shifting from 'model malice' to 'the design of system boundaries, monitoring, and communication paths.'
OpenAI Launches Law-Specific 'Astra for Law'
OpenAI has released 'Astra for Law,' a version of GPT-6 Astra specialized for the legal domain. At launch it ships with 26 partner-built plugins and 47 community-made plugins, and is being offered via Trusted Access in ChatGPT and Codex. API access is planned for a later date.
According to execution results published by evaluation firm Vals, it outperformed 'general-purpose GPT-6 Astra + web search' across every price tier on legal benchmarks. But more important than the benchmark delta is the direction of the packaging: rather than each industry doing prompt engineering from scratch, frontier models are delivered as 'domain-specific products' with maintained configurations, tools, and safe-by-default settings. As the second vertical specialization after coding, it can be read as a move to further reinforce the moat of foundation model companies.
Google Overhauls Gemini Managed Agents — Separating Secrets from the Model's Context
Google has updated Gemini's managed agents with a new Antigravity-based harness. Two APIs are particularly interesting in practice. The 'Credentials API' handles secrets through placeholders and a proxy for trusted domains, a design that keeps secrets out of the model's context. Combined with the 'Files API,' which handles artifact movement and persistent sandboxes, it amounts to groundwork for running long-lived agents safely. Google claims the revamp delivers up to 30% lower costs and a 22% improvement in cache hit rates.
The same direction is visible at other companies. Perplexity's 'Computer,' Base44's phone-calling agent, and Meta's desktop 'Muse for Mac' — commentary summed it up as resident agents with explicit permissions and user-specific context becoming the standard UX.
SoftBank Expands Arm-Collateralized Loan to $25 Billion
On the funding front, a Japanese company stood out. Sources revealed that SoftBank Group has expanded a margin loan collateralized by its Arm Holdings stake by $5 billion, bringing the total to $25 billion. After revising the terms, the company signed the agreement with creditors this month. The move appears aimed at raising capital for its ever-expanding AI investments. As funding continues to be secured through expanded credit lines, where the next investment lands remains a point of ongoing interest.
In Brief — Mozilla's '4.4-Month China Gap,' Altman's Math Ratings, Local LLMs
**Mozilla: 'China's open-weight models are 4.4 months behind the US.'** Mozilla's analysis finds that China's leading open-weight models have narrowed the capability gap with US frontier models to roughly 4.4 months. They still trail on some of the hardest benchmarks, but inference and API costs are dramatically cheaper. On Reddit, views such as 'already good enough to use' and arguments that China would have pulled ahead were it not for export controls sparked debate.
**CEO Altman: 'GPT-5.5 is an average math professor.'** In an interview at Dreamforce 2026, OpenAI's Sam Altman rated GPT-5.5 as an average math professor, 5.6 as top 1–2%, Astra as slightly above that, and an undisclosed internal model as 'able to solve problems that even the world's best mathematicians cannot solve.' He offered no specific benchmarks, and skeptical reactions were prominent on X.
**Local LLMs: ternary-quantized Bonsai 2 and Swift Qwen.** PrismML's ternary-quantized model 'Ternary Bonsai 2 (27B)' has been released at a size under 6GB. It claims to compress Qwen3.8-27B to one-ninth of its size while retaining 98.2% of its performance, but skepticism dominates the community's response. Meanwhile, 'Swift Qwen 3.8 27B,' a fine-tune that penalizes excessive thinking to cut token consumption by 58.3%, surpassed 100,000 downloads within six days of release, and reports of real-world local deployment are growing. On the research side, a paper also appeared that elucidates the mechanism of 'length inflation' — where misaligned EOS token preferences between teacher and student in on-policy distillation (OPD) cause responses to become needlessly long.