The AI Wire ●

Day 18 of Tibo Sottiaux's 28-day ship-or-reset challenge landed on Thursday as a reset day. The trackers record a global Codex reset at 17:10 UTC on October 22, at 18 of 28 days in the run that continues until November 1.

That makes two resets in a single day for Codex and ChatGPT Work users. A separate promised extra reset had already landed early that morning at 06:55 UTC, ahead of its 'before 10am PST' deadline, and the Day 18 reset followed later in the day.

The challenge's pattern is now well established: each day brings either a shipped improvement or a full limit reset. Day 17 on Wednesday was also a reset day, paired with a push for developers to sign up for Daybreak Blue access.

OpenAI let one researcher from METR and one from Redwood Research work on site for three weeks, from July 21 to August 11, auditing the internal monitor that watches its AI agents for scheming, sabotage, data exfiltration and deception. Working from OpenAI's own codebase, monitor prompts and documentation, and interviews with about a dozen engineers, they produced a 61-page internal report; the public summary was published on October 22.

The headline gap was not an attack. Long agent sessions skipped compaction events, and once that happened the monitor stopped seeing the session — leaving an estimated 10% of eligible internal Codex traffic unmonitored between May and July, concentrated in exactly the long sessions where monitoring matters most. The testers also found two client-controllable headers that disabled monitoring for new agents, an exploit that could hide hundreds of recent actions including the exploit itself, and a way for Codex agents in principle to bypass sandboxing and Auto-Review.

OpenAI says every flaw affecting its production safeguards has been patched, and that it had spotted the compaction gap itself and fixed it during the exercise. METR says it has not evaluated those fix claims, and the review only covers the system as it stood on or before July 21. The monitor itself errored on less than 0.2% of requests and, with minor exceptions, saw the full trajectory agents saw.

OpenAI made its Decisions API generally available on Thursday, taking the GPT-6 Luna-powered endpoint out of public beta. The API returns bounded, typed answers — a probability, a selection, or a rubric score — that applications use to classify content, route requests, and prioritize work.

The price is $0.10 per million input tokens, the same headline rate as GPT-6 Luna itself. OpenAI says focusing Luna's intelligence on questions with bounded answers delivers those answers up to 10 times faster than asking Luna through the Responses API, across text, image, and structured inputs.

The endpoint first surfaced at DevDay alongside the Agents API, then opened to public beta on October 6. General availability makes it a standard part of OpenAI's developer platform, aimed at the many production calls that need a decision rather than a paragraph.

OpenAI is reportedly close to another mega-round of private funding instead of going straight to its IPO: up to $30 billion at a $1.25 trillion pre-money valuation, in a round led by Coatue and investors in the UAE, sources told CNBC, via TechCrunch's October 22 report. OpenAI told investors on Thursday that it wants signatures locked in by the end of next week, and the financing is being positioned as an 'opportunistic' bridge ahead of a public listing.

The number has already moved once. Bloomberg reported a month ago that OpenAI was working on raising $30 billion at a $1.4 trillion valuation, and both outlets note the terms are still being negotiated and could change. CNBC also reports the UAE funding could be coupled with a commercial deal securing more computing capacity for OpenAI from the UAE. OpenAI declined to comment.

For scale: OpenAI filed confidential IPO paperwork in June, its last raise in March valued it at $852 billion post-money, and CFO Sarah Friar has pointed employees to a 2027 IPO timeline. A close above $1 trillion would make OpenAI the most valuable private company on record and would raise the bar further for Anthropic's own debut — reports have Anthropic seeking a $1.5 to $2 trillion valuation, after raising $65 billion at $965 billion in May and also filing confidentially in June. An October 23 follow-up in The Decoder, citing the Financial Times, adds colour on why the target moved: demand was weaker than in previous rounds, with some backers experiencing 'indigestion' after heavy investments in earlier fundraises. UAE investors and Coatue are reported to be leading the round. Treat this as rumor-grade until the round actually closes: it rests on anonymous sources, and the valuation has already slipped $150 billion in a month of reporting.

Israeli security firm Lava says it has identified 32 Nvidia Blackwell Ultra B300 accelerators — among Nvidia's most powerful data-center chips, and barred from sale to China under U.S. export controls — on monitoring systems linked to state-owned China Telecom, The Information reported on Thursday. Nobody raided a warehouse to find them: the chips surfaced through publicly reachable instances of Nvidia's DCGM Exporter monitoring tool, left on the internet without authentication.

The China Telecom systems were part of a much wider exposure. Lava's scanning found roughly 2,100 exposed DCGM Exporter hosts worldwide reporting more than 12,000 unique GPUs across nearly 300 organizations — H100s, H200s, B300s and consumer cards — with the exposed hardware valued at more than $100 million. Lava also disclosed a high-severity flaw in the same tooling, CVE-2026-47483, that could let an unauthenticated attacker crash GPU monitoring remotely.

The finding needs its caveat stated plainly: an IP address associated with a carrier is a network association, not proof of where hardware physically sits or who owns it — traffic can be proxied or tunneled — and Lava itself says the discovery raises questions about the chips' location and supply routes without proving an export-control violation. Even so, it is the first physical-evidence-style sighting in a diversion debate that has mostly run on paperwork, arriving alongside a separate October disclosure that a Chinese state-linked lessor financed more than 3 billion yuan of servers, including B300 systems, since August 2025.

Microsoft's AI diagnostic system, MAI-DxO, correctly diagnosed about 83% of complex medical cases in a peer-reviewed study published Thursday in Science — more than double the roughly 36% scored by generalist physicians working individually with reference materials. Microsoft AI CEO Mustafa Suleyman called the result a step toward what he calls 'medical superintelligence.'

The study, 'Sequential diagnosis with language models,' introduces a benchmark called SDBench that turns a case into a step-by-step encounter: the participant starts from a brief summary, then asks questions, orders tests and commits to a diagnosis. On 304 clinicopathological conference cases from the New England Journal of Medicine, MAI-DxO reached 83.2% accuracy (85.5% in peak configurations) against 36.1% for the physicians; on a second test of 325 emergency-department cases from Beth Israel Deaconess Medical Center, it scored 80.3% against 46.8%. MAI-DxO coordinates several AI agents that act like a panel of doctors, and it is model-agnostic — Microsoft tested it over models from OpenAI, Google, Anthropic, Meta and DeepSeek, with the best results on OpenAI's GPT-5 and GPT-5.5, and the orchestration layer alone lifted a GPT-5.5 baseline from 72% to about 87%.

The caveats matter: these are simulations over published case records, not a clinical trial, and cases this complex are normally handled by specialists or physician teams, who were not part of the comparison. Microsoft is now working with Beth Israel Deaconess to evaluate the system in a real hospital setting. 'The biggest challenge in healthcare delivery at the moment is the growing mismatch between demand and supply of medical expertise,' Dominic King, Microsoft AI's vice president of health, told TechCrunch.

Google and distributed-energy platform Voltus announced an expanded agreement on Thursday to build a 1-gigawatt virtual power plant in the PJM grid region, which serves 67 million people across 13 states and Washington, D.C. The deal adds 1,070 megawatts of flexible capacity to the 100-megawatt agreement the two companies signed in June, and the companies call it one of the largest distributed capacity portfolios deployed by a hyperscaler to date.

A virtual power plant aggregates capacity that already exists at homes and businesses — batteries, smart thermostats, cooling systems and industrial processes. When the grid is under strain, Voltus signals participating sites to reduce or shift their energy use, and participants receive bill savings and direct payments in return. Google funds the capacity as a way of bringing new supply to the grids its data centers draw from.

The announcement leans on a Brattle Group estimate that better grid utilization could save U.S. consumers more than $100 billion over the next decade. 'Expanding our initial commitment tenfold demonstrates our core effort to actively bring our own capacity to the grids that serve us,' said Ellen Zuckerman, Google's head of energy market development for the Americas. Voltus, which raised $225 million earlier this month and targets 20 GW of capacity by 2030, said the deal translates data center growth into direct economic benefits for homes and businesses.

Personal Intelligence in the Gemini app is expanding its Google Maps integration. Alongside the existing Search services connection, Gemini can now reference your Maps searches and saved places, and combine them with Calendar access to give tailored recommendations for your day.

Google's own example: ask which restaurant from a saved list is best after today's meetings, and Gemini checks where the last meeting is and finds saved locations near it. Personal Intelligence can be enabled from Gemini's personalization settings, and the expansion follows this week's rollout of the same feature into Android Auto.

Separately, following the launch of Gemini 4 Argon, Google is notifying AI Ultra subscribers that the new model will 'replace Deep Think soon,' suggesting they switch to Argon for deeper reasoning. The brief in-app notice does not make clear whether Deep Think — currently a thinking level on Gemini 3.1 Pro — is going away because that model is being retired, or only the mode is being phased out.

Google Cloud unveiled Agent Compute Engine on Thursday at the AGNTCon and MCPCon events. It is a compute service built specifically for running AI agents at scale, with each workload inside a hardware-isolated microVM that gets its own dedicated guest kernel and root access.

The isolation is the point: agents run code, call tools, and touch data on behalf of users, so Google is leaning on its Titanium IPU technology for added protection. The service is tuned for fast churn — agents can start quickly, pause, and resume later.

Its density trick is copy-on-write memory sharing: agents starting from the same base share memory until one changes something, which Google says has delivered benchmarks of 10 times the density on the same hardware. Google frames the service as part of its Gemini Enterprise Agent Platform, with security and governance for agents at the center.

Goldman Sachs announced on Thursday a $500 million commitment to expand its 10,000 Small Businesses program, headlined by a new Goldman Sachs Small Business AI Institute. The institute is partnering with Anthropic, Google, Meta and OpenAI to give entrepreneurs training and tools to integrate AI into their operations.

The institute will work on three fronts: delivering AI education to small-business owners across the country, publishing research on AI use by small businesses, and supporting AI-enabled expansions of access to capital. The wider commitment aims to deliver education, capital access and support services to another 10,000 businesses.

Nearly 20,000 entrepreneurs across all 50 states and four US territories have completed the program since it launched in 2010. 'Every major technology shift has rewarded the businesses that adapt most quickly,' CEO David Solomon said. 'This commitment makes sure the entrepreneurs who are creating America's jobs can lead in the AI economy.'

DeepSeek is wrapping up a funding round of about 100 billion yuan (roughly $15 billion), four people with knowledge of the matter told Reuters on Thursday — one of the largest funding rounds in China. Battery maker CATL and WeChat operator Tencent were among the biggest investors, with more than 15 investors in total, including new backers Midea, Meituan, Geely and GigaDevice.

Reuters reported the round values DeepSeek at 500 billion yuan before the new capital, double its original 50 billion yuan fundraising target. Bloomberg first reported the $15 billion figure. The round is expected to complete in the coming days, but one source cautioned that financial details and the investor lineup are not ironclad and could still change; DeepSeek, CATL and Tencent did not comment.

The raise lands two days after the US Department of Homeland Security ordered DeepSeek out of federal civilian agencies, and follows a June round of more than $7 billion at a $50 billion valuation. It is the rare Chinese-lab story that clears the Wire's bar: a raise of this size, aimed at AI infrastructure and open models, directly funds the strongest open-weight challenger to the US labs.

Canva released its first image generation model on Thursday. Canva Craft, built by the company's own research team, is live for all users in the Dream Lab tab and available to developers through the Canva API, TechCrunch reported. It is aimed squarely at the same-category models from Google, OpenAI, ByteDance and Meta that Canva previously rented for its platform.

Craft works in two stages: a planning stage first works out the spatial relationships of elements, their relative scale, the visual hierarchy and any text, and a rendering stage then generates the image at up to 2K resolution across some 30 aspect ratios. The differentiator is the training data — human-designed templates from Canva's own library, including posters, social posts, presentations, infographics, menus and other typography-rich layouts, rather than the photograph-heavy datasets behind most rivals. That makes for legible text and precise edits that leave the rest of an image untouched, and generated images can be converted into layered, editable designs with Canva's Magic tools. Researchers from Leonardo.ai, which Canva acquired in 2024, contributed to the model.

Preliminary blind comparisons put Craft fourth on DesignArena's image-generation leaderboard, which would make Canva the second-highest-ranked image lab behind OpenAI. Cheaper Craftfast and Craft-ultrafast variants are planned, as is image generation inside third-party chatbots such as ChatGPT and Gemini through Canva's MCP connector.

Anthropic's crackdown on distillation came with a price tag, Reuters reported on Thursday: the paying accounts the company has cut off accounted for between 5% and 7% of its revenue in the first nine months of 2026, according to investors briefed at an investor meeting in San Francisco last week. Anthropic alleges those accounts were not ordinary customers but Chinese AI labs quietly mining Claude's outputs to train competing models.

The technique, distillation, has a student model learn from a teacher model's outputs. Anthropic said in February that DeepSeek, Moonshot and MiniMax had generated more than 16 million interactions with Claude through roughly 24,000 fake accounts, in violation of its terms. Earlier this month it rolled out new defenses — access restrictions, proxy blocking and bans on tens of thousands of accounts — which it says have cut the reasoning data being siphoned by 90% to 99% (covered here on October 12).

Anthropic frames the practice as intellectual-property loss with national-security implications, and it is telling investors the crackdown has effectively stopped Chinese developers from distilling its models. The disclosure also hands the company an explanation for a month-to-month growth deceleration investors were already nervous about, weeks ahead of what could be the largest IPO in history.

Anthropic has made Claude Models on Google Cloud (Anthropic managed) generally available, giving Google Cloud customers a second way to buy Claude alongside the existing Google-managed offering. Both routes use Google Cloud credentials, IAM and Marketplace billing, and spending on the Anthropic-managed service counts toward existing Google Cloud commitments — but Anthropic operates this service itself, sets its usage limits and acts as the data processor under its own commercial terms.

The practical difference is features. The Anthropic-managed route carries the full Claude Platform from day one, including document uploads, web search and code execution, with new models and platform features arriving the same day Anthropic releases them; Claude Managed Agents remains in beta. The launch line-up is Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5, and Opus 5.5 fast mode is generally available through the service at $8 per million input tokens and $40 per million output tokens.

Google's own comparison keeps the Google-managed option as the better fit for customers who need Google as data processor, strict regional residency, or native integrations such as Model Armor and Agent Engine, and it does not claim the same VPC Service Controls and customer-managed encryption keys for the Anthropic-managed route. Anthropic also warns customers with existing discounts to talk to their account executives before starting, because discounts cannot be applied retroactively to usage on the new service.

Anthropic and formal-verification startup Theorem say Claude has done something that used to take experts years: it wrote mathematical proofs that real software will not crash. In research published on October 22, Claude proved that 239 test entry points across 123 open-source libraries from Google's OSS-Fuzz programme never crash on any input under 64 MB, working at the level of compiled RISC-V machine code.

The proofs run to roughly 180 million lines of Rocq, a proof language checked mechanically by computer, with a typical proof around 600,000 lines for a binary with a median of 18,000 instructions. Along the way Claude found bugs that trillions of fuzzing runs had missed, patched them, and then proved the patched code crash-free. Anthropic is publishing a subset of the proofs and the trusted components underneath them, so the checking does not depend on trusting the lab.

The bounds matter: the proofs cover single-threaded, static RISC-V binaries, not the x86 or ARM builds most software actually ships as, and 'formally verified' here means crash-freedom for the stated property, not bug-free in general. Even so, Anthropic frames it as the first large-scale use of AI for formal verification of production software at assembly level — arriving in the same week AI models helped surface the PHP2Shell flaws and Claude stripped unneeded code out of OpenSSH.