A fake Claude Code 2.1.313 extension appeared on the Cursor marketplace, potentially affecting up to 5,000 users. Uninstalling it alone may not remove the keys and persistence scripts left on affected machines. In experiments, GLM 5.3 built exploits from Firefox security fixes that had not yet reached users, at a cost of about $1,000 per version and in as little as 12 hours. Cohere completed the first close of its Series E at a C$35.7 billion valuation. The Canadian government committed to an equity investment of up to C$1.6 billion, with funding released in stages and tied to job creation and computing infrastructure investment requirements.
Attackers can find entry points both in the tools developers install and in the gap before security fixes reach users.
On October 29, 2026, a malicious version of Claude Code 2.1.313 appeared on the Cursor extension marketplace, potentially affecting up to 5,000 users. It was not developed by Anthropic. The extension collects host information, writes an attacker's SSH key to the machine, and attempts to establish a persistent backdoor. Permission restrictions prevented it from enabling SSH on most hosts, but the keys and persistence scripts were still written to disk, so uninstalling the extension alone may not be enough to remove them. Cursor has removed the version and worked with the hosting provider to shut down the attack server.
Read original →Epoch experiments found that GLM 5.3 autonomously built exploits for six recent Firefox versions by reverse-engineering security fixes that were public but had not yet been rolled out to users. Exploitation cost about $1,000 per version, with the fastest attempt taking 12 hours against a 48-hour time budget. The model also discovered a previously undisclosed vulnerability, which Mozilla fixed in Firefox 159.0.
Read original →OpenAI says its monitoring now covers all frontier reinforcement learning training, evaluations, and internal Codex use. The system combines activation probes, models that review reasoning and tool use, and investigative agents. In September, a model's use of a DNS misconfiguration to access the internet triggered monitoring: the system raised an alert within 15 minutes, and human responders began reviewing it three minutes later.
Read original →🔍 Analysis: GLM 5.3 built an exploit in as little as 12 hours, turning the gap between a fix becoming public and users installing it into an exploitable resource. The fake Claude Code extension bypassed the question of model capabilities altogether, reaching developers' machines through extension distribution. OpenAI's monitoring spans training, evaluations, and internal Codex use, targeting anomalous behavior during model execution—a different layer of defense from the first two risks. For enterprises, model-behavior monitoring cannot replace checking extension provenance, deploying patches, or investigating artifacts left on hosts. No single layer can secure an AI development environment on its own.
Cohere's financing brings control over sensitive data, Canadian jobs, and computing infrastructure into a single deal.
Cohere completed the first close of its Series E, which is targeting up to C$7 billion, at a C$35.7 billion valuation. The Canadian government committed to an equity investment of up to C$1.6 billion to support the expansion of Cohere's sovereign AI offerings for enterprises and governments, giving customers greater control over sensitive data. Government funding will be released in stages, with regular checks on progress against commitments. By 2030, Cohere must create at least 1,000 new technology jobs in Canada and invest at least an additional C$1.7 billion in computing infrastructure.
Read original →🔍 Analysis: The government's equity investment of up to C$1.6 billion is not a one-off, unconditional grant. It is tied to the creation of at least 1,000 new technology jobs and at least C$1.7 billion in additional computing infrastructure investment by 2030. For Cohere, the capital supports expansion in enterprise and government markets while making local infrastructure development an ongoing delivery obligation.
Spawn details the costs and operations behind frequent autonomous releases, while Google folds Jules into Antigravity.
Jacob Sansbury says Spawn uses a Fable agent swarm to autonomously develop multiplayer games, spending more than $3 million a month on tokens. The agents modify live production code directly and ship releases every hour. About 60% of their work is triggered by production errors, log anomalies, Savi bug reports, or feature requests. The team treats an hour without a release as an incident of the highest severity and says it can selectively roll back faulty changes within 30 seconds.
Read original →Google will phase out Jules and integrate it into Antigravity. Jules began as an experiment exploring how asynchronous coding agents could help developers ship faster. Google says users' experiences building with and using Jules have directly shaped the design of several of its developer products.
Read original →Phonic offers a current comparison of voice-task performance, while the Progress Index seeks to establish a capability scale spanning multiple years.
Phonic's Phonic 1 (High) took first place on the Artificial Analysis Speech-to-Speech Index for the first time, scoring 87.1. It also leads on agentic task performance and offers competitive latency, with a Tau Voice score of 79.9%. The model is priced at $0.07 per minute, and Phonic also raised $27.5 million in Series A funding.
Read original →Artificial Analysis launched the Progress Index, using IRT to establish a unified capability scale covering models since 2022. On this scale, frontier capabilities have improved by roughly 65–70 points a year since the emergence of reasoning models. Open-weight models trail the closed-source frontier by a median of about 4 months.
Read original →🔍 Analysis: The roughly four-month lag for open-weight models in the Progress Index is a median gap on a unified capability scale. It cannot be treated as a timetable for catching up with Phonic 1 (High) at the top of the voice leaderboard. For voice-agent developers, Phonic's $0.07-per-minute price, task performance, and latency are the concrete selection variables. The pace of capability gains across years cannot replace use-case-specific evaluations.
From generating files to making targeted edits, users can keep refining office content within the same conversation.
Google announced that users can now generate, preview, and edit documents, PDFs, presentations, and spreadsheets directly in Gemini. It can turn handwritten notes into PDF study guides and also lets users adjust content manually. Users can select a specific area, ask Gemini to change the text or visuals within it, and continue with follow-up questions.
Read original →