01Dario Amodei Asks the Industry to Slow Down — and Musk and Altman Say Yes [AI Applications & Industry]Anthropic's CEO published 'We Must Pace the Frontier,' arguing labs should deliberately slow capability gains so alignment, security and third-party evaluation can catch up — warning that an unchecked agent swarm could take over 'the entire internet' within 6-12 months. Anthropic is giving evaluators like METR permanent, employee-like access; within hours Musk posted 'Dario is right' and Altman committed OpenAI to evaluator access, while Hugging Face asked for a seat for open source. → source
02Hundreds of AI Agents Breached 395 Organizations in One Campaign — 11 in 26 Seconds at Peak [AI Applications & Industry]GreyNoise researchers traced a Russian-speaking threat actor that ran hundreds of AI agents built on OpenAI Codex and a DeepSeek model to exploit two PaperCut NG/MF flaws — 440 instances at 395 organizations across 48 countries. First remote code execution came in under four hours; credentials were harvested from 280 organizations, with education the hardest-hit sector. → source
03Cohere Nears a $2-3B Raise at $20B — With Two Governments Writing Checks [AI Applications & Industry]The enterprise-LLM firm is in advanced talks to raise $2-3B at a $20B valuation — nearly triple last year's $7B — in what would be the largest private financing on record for a Canadian startup. Ottawa is set to participate alongside Radical Ventures, Inovia and PSP, a notable twist as Berlin-backed Mistral and Washington-tied Anthropic each anchor their own sovereignty play. → source
04Sam Altman: It Would Be 'Ill-Advised' for OpenAI to Go Public in 2026 [AI Applications & Industry]OpenAI's CEO told TechCrunch that going public next year would be ill-advised, pointing to the strangeness of an IPO while the company is spending at warp speed. In the same week he hinted at a coordinated industry pact on safety — while Anthropic's own IPO looms as the market's test case for AI disclosures. → source
05Salesforce Ships Seven Named Agentforce Agents — and Counts 7 Billion Work Units [AI Applications & Industry]Casey, Paige, Carter, Hunter, Marshall, Piper and Fin arrive as job-ready agents built on a new long-horizon runtime with persistent memory and durable execution — all generally available except outbound-sales Hunter, which pilots in November. Salesforce says it has delivered 7 billion Agentic Work Units to date and cites Fin autonomously resolving 79% of Anthropic's own support conversations. → source
06Agnes-3.0-Flash: A 33B Open-Weight Multimodal Model with 262K Context, Apache 2.0 [AI Models & Research]Agnes-AI released a 33B-parameter multimodal model under Apache 2.0 with a 262,144-token context and a hybrid architecture mixing 54 delta-rule recurrent layers with 18 global-attention layers. It posts 85.05 on GPQA Diamond and fits on a single H200 at bf16 — a serious open-weight multitasker you can actually run. → source
07Sakana's Fugu Max Turns a Pool of Open Models into One API — and Outscores Frontier on Six of Ten Benchmarks [AI Models & Research]Sakana AI shipped Fugu Max and Fugu Ultra v2, a learned orchestrator that routes each query across open-weight and specialist models behind a single OpenAI-compatible API. Fugu Max ranks best on six of ten benchmarks including Terminal Bench 2.1 — at $2/$6 per million tokens, 40-60% below Sonnet 5 and GPT-5.6 Terra, without any frontier model in its pool. → source
08Grok 4.7 Delayed Again — Musk Says RL Taught It to Give Up on Solvable Tasks [AI Models & Research]xAI pushed the 2.1-trillion-parameter Grok 4.7 past its Sept 12 target after Musk wrote that the team 'might have penalized response length too much in RL, as it still gives up on hard tasks (that it can do!) too early.' It's the second delay for the model, which is trained partly on SpaceX's own rocket and manufacturing data. → source
09SGLang Flaw Becomes the Fourth Critical AI-Inference CVE in Four Weeks [AI Tools & Ecosystem]CVE-2026-86793 lets an unauthenticated attacker execute code through SGLang's /update_weights_from_tensor endpoint when no API key is configured — and maintainers acknowledged the bug in July but shipped no patch before disclosure. It's the fourth critical flaw in AI inference infrastructure in four weeks, after Ollama, DeepSeek Harness and IBM Langflow. → source
10Claude Code, Codex and Cursor Sandboxes Leak — Anthropic Took 50 Days to Patch [AI Tools & Ecosystem]Stealth startup Accomplish disclosed leaky sandbox vulnerabilities across Claude Code, OpenAI Codex and Cursor after quietly reporting them this summer. Cursor and OpenAI fixed their bugs in about a week; Anthropic needed roughly 50 days and 30 releases — prompting its CTO to say security talk 'doesn't really reflect in how they actually build products.' → source
11Minitap Accuses Google of Erasing Engineers from an Open-Source Mobile-Agent Repo [AI Tools & Ecosystem]Paris-based Minitap says Google's Pixel test-engineering team copied code from its Apache 2.0 mobile-use repo into the ARTEMIS agent, then force-pushed the three engineers' names into an unrelated author identity — and left Minitap's own 100% AndroidWorld result off the leaderboard. Google opened a credit-restoring PR only after a public issue hit #3 on Hacker News. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.