LastWeekIn.Tech

A quieter news day dominated by model and safety updates from the major labs: OpenAI touts a wave of AI-generated math proofs amid bubble-burst warnings, Anthropic expands red-team access while reporting a huge vulnerability haul, a new adversarial-poetry jailbreak technique surfaces in the wild, and incremental model releases continue from Anthropic and the open-source community.

1. 1

   Number 1. Capabilities foxnews.com

   OpenAI publishes 722 AI-generated mathematical manuscripts containing new proofs

   OpenAI published proofs for 722 previously unsolved mathematical problems in a public GitHub repository, generated by an unreleased internal model that was handed roughly 4,000 problems, averaging about three hours of ChatGPT Pro compute per result. Sam Altman shared the repository publicly, while investor Ray Dalio separately warned the AI industry bubble may be nearing a burst.

   Why it matters If verified, this represents one of the largest batches of AI-generated mathematical proofs released to date, suggesting frontier models are becoming genuinely useful research collaborators rather than just chat assistants. The scale (722 proofs from a single prompt-driven agent run) also raises questions about verification rigor and whether external mathematicians will validate the claims, especially as skepticism about AI capability claims grows alongside bubble warnings from figures like Dalio.

   What to watch Watch for independent mathematician review of the proofs' validity and whether OpenAI releases the underlying model publicly.

2. 2

   Number 2. Security & Safety The Hacker News

   Anthropic Expands Claude Access for Vetted Cyber Teams as Glasswing Finds 129,000 Flaws

   Anthropic is expanding its program allowing vetted cybersecurity professionals to test Claude with reduced safeguards, reporting that its Project Glasswing initiative found at least 129,000 verified software vulnerabilities between April and July 2026, plus 5,500 more through open-source scanning through October. More than 33,000 of the verified vulnerabilities were rated critical or high severity.

   Why it matters This is a significant demonstration of AI-assisted vulnerability discovery at scale, showing both the defensive promise and dual-use risk of frontier models with loosened guardrails. The sheer volume of critical findings suggests AI-driven scanning could reshape bug bounty economics and patch cycles industry-wide, while also underscoring why access controls on 'reduced safeguard' models matter.

   What to watch Watch whether other labs launch similar vetted-researcher programs and how disclosure coordination scales with this volume of findings.

3. 3

   Number 3. Security & Safety The Register

   PoeLLM malware infects exposed AI servers in cryptomining attacks

   Security researchers at Lumen's Black Lotus Labs identified PoeLLM, malware that has infected more than 3,000 servers since April by using a poem-based jailbreak to trick AI models into bypassing safety guardrails, in what researchers call the first real-world case of 'adversarial poetry' jailbreaking. The malware breaks into enterprise AI infrastructure to mine cryptocurrency and expand a botnet, with a suspected Italian attacker behind it.

   Why it matters This marks the jump of a previously theoretical AI jailbreak technique—reformatting harmful prompts as poetry—into an active, widespread criminal campaign against AI infrastructure. It signals that adversarial prompting techniques are maturing from academic red-team exercises into operational attack tooling, raising the bar for how AI safety classifiers need to handle stylistic obfuscation.

   What to watch Watch for other adversarial-poetry or style-based jailbreak variants and whether AI vendors push targeted patches to their safety classifiers in response.

4. 4

   Number 4. Companies Hacker News

   Claude Haiku 5.5

   Anthropic released Claude Haiku 5.5, described as its cheapest, fastest, and most capable small model yet, aimed at high-volume tasks like summarization, classification, and live customer support. The company said it costs substantially less than Haiku 4.5 and is designed to pair with Opus 5.5 and Sonnet 5.5 as a subagent for coding workflows.

   Why it matters Small, cheap, fast models are increasingly the backbone of agentic systems that rely on subagents for routine tasks, and Anthropic's aggressive pricing on Haiku 5.5 intensifies competition at the low-cost end of the model stack where margins and volume matter most. This also reflects the industry's broader shift toward multi-model agent architectures rather than single large models for every task.

   What to watch Watch for competitive responses from OpenAI and Google on small-model pricing, and adoption data on Haiku 5.5 as a subagent in production agent pipelines.

5. 5

   Number 5. Open Source Hacker News

   Strands Decider 2B: a small, open-source, decision model

   A new open-source small decision model, Strands Decider 2B, was released as part of the strands-labs initiative for hands-on agentic AI experimentation. Unlike general-purpose LLMs, decision models are designed specifically to select among fixed sets of options rather than generate arbitrary text, following the emerging 'system one' model category popularized by TypeSafe AI's Jev.

   Why it matters This points to a growing architectural trend of separating fast, constrained 'decision' models from larger generative LLMs within agent pipelines, mirroring dual-process cognitive theory. Open, small decision models could become a standard low-latency component in agentic systems, reducing cost and improving reliability for routing and classification tasks that don't need full generative capability.

   What to watch Watch for benchmark comparisons between dedicated decision models and prompted LLM classifiers, and whether major agent frameworks adopt this decision-model pattern broadly.