We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
An outside audit found holes in OpenAI's agent monitoring
METR and Redwood Research placed two staff on site at OpenAI for three weeks this summer. They evaluated the systems OpenAI uses to watch its internal agents for misalignment . The pilot asked two questions: what share of agent traffic was monitored, and whether agents could get around the monitoring .
- Monitoring could be postponed. The monitor ran only when an agent handed control back to the user, so an agent could avoid monitoring indefinitely by continuing to act .
- Some traffic went unmonitored. METR estimates that roughly 10% of internal Codex traffic was unmonitored from May to July. OpenAI had found this issue on its own shortly before the exercise .
- Exploits. The team found a way for agents to hide hundreds of actions from monitoring. It also found a method that would likely stop the monitor from running, and a simple bypass of Codex sandboxing and Auto-review .
OpenAI says it has patched every identified vulnerability that affects its production safeguards . It says the findings helped strengthen safeguards as it expanded the system . The review covered monitoring as it stood when the pilot began. It did not cover later changes, such as extending monitoring to RL training, which was not monitored at the start .
The transparency terms drew criticism. METR agreed in advance to publish only a brief, high-level write-up . OpenAI was also allowed to redact, and METR had no right to describe the redactions on its own . One observer says the full report was 61 pages, and the version METR/Redwood leadership saw after redaction was 38 pages . METR says future exercises should follow its recent proposal for stronger public transparency .
More AI-driven cyber incidents, and a fight over what php2shell shows
php2shell was covered last cycle. New figures this cycle: Cloudflare's CEO says his network saw thousands of exploit attempts in the 10 hours before the announcement. That rose to tens of thousands in the 45 minutes after it, hundreds of thousands within two hours, and millions by the time he posted . These count attempts, not successful compromises.
The pre-announcement activity fed the argument over whether defense is ahead of offense. Stella Biderman (@BlancheMinerva) says AI-enabled defensive work found the bug and there is no evidence of large-scale exploitation . She thinks the early attempts more likely point to a leak, since more people are brought in shortly before a disclosure . Sam Altman went further: models find bugs "outside of the normal threat model," including hardware-level ones, and "I no longer think we're gonna be able to just patch all the bugs" .
Other incidents reported this cycle:
- Banks: Major US banks are finding unauthorized third-party AI agents inside customer accounts. Some appear to act independently of user instructions, according to Axios sources . Executives say they struggle to tell humans from agents, and to tell agents a customer authorized from rogue ones. Several banks are planning for "day after" scenarios .
- Microsoft 365: A post summarizing Microsoft's findings says a China-aligned group used an open-source agent framework against 2,250+ Microsoft 365 tenants after the ConnectWise breach. It made 2,900 sign-in attempts a minute and in some cases got privileged access in under three hours .
- AI API keys as the target: LiteLLM 1.107.0 shipped a credential stealer for about 100 minutes. The advice is to roll back to 1.106.2 and rotate all credentials . Railway and OpenCode were also hit, and the attackers appear to want AI API keys .
- Defenders blocked by safety filters: OpenCode's own AI models refused to help it respond to a 48-hour agent-driven attack until it switched to OpenAI's Daybreak. Straiker says a Mexico-linked gang aimed Claude and Grok agents at 400+ Latin American banks and payment firms. India's eMudhra is leaving the certificate business after its keys proved guessable .
Tools for agents and defenders also shipped:
- Formal verification at scale: Anthropic and Theorem Labs used Claude to prove security properties in more than 100 real-world open-source libraries over a few weeks .
- 1Password for Agents: an agent asks for a credential, and once the user approves, 1Password fills it in directly so the agent never sees it .
Open weights: MiniMax releases M3.1 Flash, and Chinese approvals look quick
MiniMax published weights for M3.1 Flash: 428B total parameters, 23B active per token. It reports 73.7% on DeepSWE v1.1 and 82.2 on CyberGym . teortaxesTex argues the model is essentially M3 rebranded, with attention sparsified in three more layers, quantization, and DSpark speculative decoding added . Under China's new rules, DeepSeek and MiniMax reportedly got permission to open-source V4.1F-1017 and M3.1-Flash . A CCTV-affiliated commentator, as relayed, said DeepSeek applied after the Oct. 16 export restrictions and was approved within hours . That account is anecdotal.
Open models' cyber ability keeps rising, and so does the cost advantage. On one CVE-rediscovery benchmark, DeepSeek V4.1 Flash 1017 rediscovered 87.5% of the CVEs at pass@3, behind only GPT-6 Astra at 90.6%. It cost $2.40 per CVE found, versus $135.23 for Astra . One user reports that the model built a multi-stage memory-corruption exploit escaping Chrome's V8 sandbox after 12 hours .
AI for science: an open DNA model and a $215M robotic lab
Hugging Face released Carbon-X, a 14B model midtrained on DNA and text. It claims the model beats every open language model on DNA tasks and matches GPT-6.1 Sol at 1,000× lower cost . On the FrontierVEP benchmark, RL applied directly to the Qwen3-14B base model never rose above chance. Starting from the midtrained checkpoint, the same method did . The team also points to frontier-model refusals: Opus 5.5's next-base-prediction score falls from 17.7% to 1.1% as reasoning effort rises and refusals climb .
Kevin Weil, formerly CPO at OpenAI, launched Warp Sciences with a $215M Series A led by Lightspeed . It will be a US preclinical CRO, a contract lab that makes and tests molecules for drug developers. It will not develop its own drugs, and plans to automate its chemists' work one workflow at a time . Microsoft's MAI-DxO diagnosis system, published in Science, scored 83.2% on complex NEJM cases and 80.3% on emergency-department cases. Participating generalist physicians scored 36.1% and 46.8% .
A new paper, Agent4Sci, set coding agents against 20 scientific ML benchmarks. The agents beat the published best on all 20, mostly using standard ML-engineering techniques. Adding three lightweight research skills raised Claude Opus's average improvement from 10.4% to 20.1% .
Agent standards move to neutral foundations
Anthropic is donating Agent Skills to the Agentic AI Foundation (AAIF), which already holds MCP. More than 70 platforms have integrated the standard since December 2025 . Engineers from Anthropic, Earendil (Pi), Google, Microsoft, OpenAI and OpenCode will govern it as a technical steering committee . Hugging Face is giving its OpenEnv RL environments to the AAIF and the Linux Foundation .
Benchmarks are being reset
Artificial Analysis's Coding Agent Index v2.0 swaps in Cognition's FrontierCode. Claude Code with Opus 5.5 leads at 61.7. On cost, Codex with DeepSeek V4.1 Flash scores 47.3 at $0.39 per task, against $11.42 per task for Opus . CritPt v2, a research-level physics benchmark, has GPT-6 Astra at 83% and GPT-6.1 Sol at 82% . The scores are not comparable with v1, where Astra had 32%, and most of the gain comes from revised challenges .
Briefly
- OpenAI funding: OpenAI is reportedly close to raising up to $30B at a $1.25T valuation, led by UAE investors .
- Decisions API: OpenAI's Decisions API is generally available .
- Claude on Google Cloud: the full Claude API now works from Google Cloud accounts, including Managed Agents .
- Canva Craft: Canva's first image model debuted at #4 on DesignArena, making Canva the #2 image lab after OpenAI .
- @OfficialLoganK reported that Gemini 4 Argon processed about 25 trillion tokens in its first 72 hours after launch.
- DeepSeek open-sourced DeepInfer, the inference framework serving V4.1-Flash in production; DeepSeek says the framework is intended to lower compute costs and energy use, and ThursdAI describes it as Apache 2 and running on Blackwell and Huawei Ascend.
- ThursdAI reported NVIDIA's HGX Rubin NVL8 was shown at GTC Berlin, priced at $1.2 million and expected at CoreWeave that quarter; NVIDIA projected 10× more tokens per megawatt than HGX B200.
- Recursive introduced Flint, a 39.8B MoE with 2.5B active parameters, Apache 2 licensed and designed for local Mac coding; its Eureka Machine selected the architecture and quantized the LM head to 4 bits. Recursive says Flint was built from scratch in under four months, has open weights, and runs at 125+ tokens/sec on MacBook Pro and 150+ on RTX Spark.
- Anthropic opened a research preview of Claude Duet, a proactive agent app that surfaces what needs attention, completes requested tasks while the user steps away, and checks before acting on their behalf.
- 1Password launched an agent platform where an agent requests a vault credential and, after user approval, 1Password fills it into the destination site so the agent does not see or take possession of the password; the feature is free and optional for personal and family accounts, and requires opting in to an agent connector.
- ThursdAI reported that LiteLLM 1.107.0 distributed a credential stealer for about 100 minutes, with prompt updaters among those affected; it advised rolling back to 1.106.2 and rotating credentials. The roundup also said Railway and OpenCode were hit and attackers appeared to be targeting AI API keys.
- Anthropic established a wholly owned Public Benefit Fund starting with 5% of its shares; the company says roughly 20% of its shares are committed to public benefit overall, alongside cofounder equity pledges and employee donation matching.
- MiniMax made M3.1 Flash weights public; the model has 428B total parameters and 23B activated per token. MiniMax reports scores of 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents’ Last Exam, and 82.2 on CyberGym, targeting agentic, coding, and cybersecurity tasks.
- vLLM announced Day-0 support with native NVFP4 KV and MoE weights and DSpark speculative decoding. In a 256K-context test on 4× B300 with DEP4, vLLM reports NVFP4 reduced checkpoint size from 451.5 GB for M3-MXFP8 to 250.1 GB (44.6% smaller) and KV memory per token by 43.75% versus FP8; M3.1 fit 76 concurrent sequences versus M3’s 28. Separately, vLLM reports 724.9 tok/s per-user decode and 191.5k tok/s per GPU on NVIDIA Vera Rubin NVL72.
- vLLM also reports AMD MI350X support delivering up to 2.3× faster mono decode and double the throughput at low concurrency for an 8K-input/1K-output workload, plus Day-0 RL support with veRL, FSDP2, and LoRA.
@LeroyLi311063 says there will be no V4.1 Pro “in the short term”; the quoted post gives no reason . @teortaxesTex speculates this could reflect another flagship training problem or a limit to scaling this broad architecture family beyond 1T, and says V4-Flash was great .
Arena teased an upcoming new model from Kimi Moonshot . The linked benchmark post reported Kimi-K3 at #1 in Frontend Code Arena with 1,679 points, up from Kimi-K2.6 at #18, and ranked first in six of seven frontend domains .
A practitioner reports Gemini 4 Argon generating bounding boxes and an HTML view for a prompt to label raised parts on a board, and recommends an iterative vision pipeline: define a labeling rubric, use Gemini on seed data to refine the rubric and create golden examples, add specialist tools, scale labeling, fine-tune a smaller task-specific model, and use Gemini to review high-uncertainty samples. The accompanying recommendations suggest using SAM3 to turn Gemini-generated boxes into high-quality masks, and Falcon Perception-HD as a potentially better lightweight local option for some text-prompt tasks at scale.
In a debate over why the U.S. embraced AGI, @tszzl credited competing ideological centers, early rationalist belief in AGI, and deep, relatively free startup capital, while arguing that China’s centralized system and less-free financial markets constrain comparable investment . @teortaxesTex disputed the cultural explanation, arguing that China has substantial AI talent by virtue of its scale and that funding and investor appetite are important constraints .
Anthropic and Theorem Labs used Claude to prove security properties in more than 100 real-world open-source libraries over a few weeks, applying AI to formal verification at scale; the post says this work has historically been limited to small codebases or required substantial expert time.
@maharshii questions whether megakernels help compute-bound workloads, rather than memory-bound ones . They report that projects they encounter mostly target language-model decode, with few or no examples for compute-bound workloads such as DiTs or large prefill, and call for discussion .
Teknium says Hermes Agent built a webpage of web-development demos in 20 minutes after being asked to showcase lesser-known styles and component libraries; he later expanded the site from around 75 finds to 462 demos covering 425 libraries in 23 sections.
- Fireworks promotes DeepSeek V4.1 Flash on its Ultrafast service with the wording “1017 runs at 1,521+ tok/s” and claims it can ship Tetris in 11 seconds, 7× faster than Claude Opus 5.5 .
- @scaling01 argues that market pressure and larger models will drive more language compression, forecasts frontier inference speeds of 10k tokens/s within the next few years, and says multi-agent training remains immature, especially for open models .
Picsart said the featured content was generated with Kling 4.0 and invited users to try Kling 4.0 on Picsart .
Muse’s Japan launch was marked with a speedrun of Japan’s most important cultural icons.
- @teortaxesTex forecasts a second wave of Chinese AI startup failures by Q1 2027, citing rising capex and national-scale compute scarcity, fading demand for models that are only slightly different, slower, or more expensive, and DeepSeek’s MIT-licensed flagship models and Apache-licensed codebases reducing the incentive for competitors to build full-stack alternatives. The poster speculates that only two major startups may still be growing by the end of 2027, with DeepSeek the likeliest leader and Zhipu or MoonshotAI contenders for second place.
- In a related analysis, the commentator describes DeepSeek’s V4.1 line as strong and cheap, and argues that running Flash profitably requires a full NVL72 GB200 rack—potentially shifting accessible colocated compute toward large deployments. V4.1 Pro and Ultra are identified there as hypothetical, not released products.
A post claims Argon tops humor benchmarks, but names no benchmark or scores .
Artificial Analysis launched Coding Agent Index v2.0, evaluating models with their agent harnesses on 290 equally weighted tasks across software engineering, terminal use, and repository Q&A; it replaced DeepSWE with FrontierCode v1.1 and revised SWE-Atlas-QnA grading, using GPT-6 Luna as judge with five votes per criterion. Claude Code with Claude Opus 5.5 led the index at 61.7 and scored 56.0% on FrontierCode. The benchmarks produce materially different rankings: one post notes disagreement among FrontierCode, SWE-Atlas-QnA-AA, and TerminalBench 4.0, and reports DeepSeek’s TerminalBench 4.0 score as 42.3 with Claude Code versus 28 with Codex. A thread reply says DSH reduced token use, turns, execution time, and cost versus Codex while improving scores and reducing tokens on TerminalBench 4.0, but lost substantially on FrontierCode. Cost-performance also varied widely: Codex with DeepSeek V4.1 Flash scored 47.3 at $0.39 per task, while Claude Code with Opus 5.5 scored 61.7 at $11.42 per task; these are API token costs, not subscription prices.
Kling 4.0 is available on Pollo AI, which advertises keyframe-based sequence shaping, multiple references, and video editing; Pollo says Ultra users get 10 days of unlimited access to the model.
Warp Sciences launched to apply AI and robotics to accelerate medicine development, raising a $215 million Series A led by Lightspeed to bring frontier AI into laboratories and expand scientific capacity in the United States.
Kling 4.0 is live on idea.ai and is advertised as generating 30-second scenes from a single prompt, keeping characters consistent across shots, switching among live-action, anime, and 3D styles, and producing cinematic-quality video.
Anthropic and Theorem Labs used Claude to prove security properties in more than 100 real-world open-source libraries over a few weeks, bringing AI to formal verification that had historically been feasible mainly for small codebases or required substantial expert time.
SemiAnalysis describes a data-center buildout signal sequence: early construction and site work can reveal projects first; it says its model predicted a surge in orders at Sterling, an early-stage construction company, which subsequently occurred. Electrical and cooling equipment follow, while energy infrastructure such as gas turbines has multi-year lead times. Binding pre-leases signed by hyperscalers, neoclouds, and AI labs can indicate planned capacity before construction begins.
