Uncategorized

DailyAI Briefing: Nvidia board probes Huang, plus 4 more

Welcome to the DailyAI Briefing, the day’s five biggest AI stories. Nvidia shares fell 5% as the board opened a probe into whether Jensen Huang was targeted by foreign intelligence. An AI agent copied Azure Key Vault data in five minutes, DeepSeek’s Harness hit 1.54 million daily users, the best coding agent fully cleared only 26.4% of real tasks, and from Thursday Claude-driven machines must stop if the human link drops.

Let’s dig in.

Nvidia’s board probes Huang after an affair

5%. That is how far Nvidia shares fell on Monday, CNBC reported, after the board said it would investigate whether chief executive Jensen Huang was the target of a foreign intelligence effort. The inquiry was announced on Sunday, shortly after Huang told staff he had been unfaithful to his wife.

“I was unfaithful to my wife. My actions hurt her and our children. There is no excuse for what I did. I am deeply sorry,” he wrote. The New York Times and the Financial Times identify the woman as Alice Huang, also reported as Huang Le, a Chinese citizen linked to Megaspeed, a Singapore cloud firm that buys Nvidia chips. They are not related. Those reports say US officials concluded she was a Chinese intelligence asset; Reuters said it could not independently verify the claims.

Independent directors, with outside counsel, will run the review. The board said it has “no reason to believe that either national security or NVIDIA secrets were compromised,” in a statement carried by CNBC, and Huang stays in post. China called the allegations “pure groundless speculation and slander,” foreign ministry spokesperson Guo Jiakun said, according to ABC News. Trade records cited by the Times show a Malaysian Megaspeed affiliate bought at least $4.8 billion of AI equipment containing Nvidia technology between June 2024 and February 2026. Three Republican House chairs have asked Huang for a closed-door briefing on 18 November, a day after earnings. We set out the inquiry here.

An AI agent emptied a Key Vault in five minutes

Five minutes. That is how long it took an AI agent to go from scanning public GitHub for leaked secrets to copying data out of an Azure Key Vault, Microsoft Threat Intelligence said in a 5 November report. Attackers used an agentic offensive framework called ARTEX to hunt credentials, application logins and AI API keys, then use them against Azure and Microsoft Entra.

The scans began as early as August 2026. From about 28 October, Microsoft saw leaked but still-valid application credentials used to sign in across multiple tenants, including vendor-owned logins that opened the door to downstream customers. Deleting a secret from a repository does not revoke it. Microsoft says organisations should treat anything found in public code as already compromised. We wrote up the five-minute case here. The leaked secrets were a doorway into Azure itself.

DeepSeek Harness hits 1.54 million daily users

1.54 million. That is how many daily active users DeepSeek says its open-source agent app Harness had as of 4 November, counting people on the company’s official API. Those users burned through about 92 trillion tokens a day, the team said in an 8 November X article. Divided out, that is roughly 60 million tokens per active user per day, far more than a normal chat session.

Harness v0.1 went live on 13 August, with a desktop app on 29 September. DeepSeek counts roughly 35,000 third-party plugins, and about 60% of active users have built or installed at least one. People running Harness on other model providers are not in the 1.54 million figure. A marketplace and token grants for plugin authors are planned, with no date yet. We ran through the company’s numbers here.

The best coding agent fully clears only 26.4% of real tasks

26.4%. That is the share of real user app-edit tasks that the strongest coding agent fully cleared on Alibaba’s Co-Design Bench, according to a paper from Alibaba Token Hub and university co-authors. The same run met 69.9% of each task’s hidden acceptance criteria on average. Most jobs were mostly done and still unfinished.

The benchmark at codesignbench.github.io builds 273 tasks from 26,314 real coding sessions. OpenAI’s GPT-6 Astra led ten model endpoints under a shared Claude Code harness. It fully passed 41.0% of 100 multimodal editing tasks, but only 17.9% of 173 repository edits that stretch across components, routes and state. Older tests are saturating: Claude Opus 4.8 reached 99.4% with agentic execution on a 150-task WebCompass-edit split. We unpacked the paper here.

Claude-driven machines must stop if the link drops

From Thursday, Claude can no longer drive injury-capable machines without a human who can watch and stop them. Anthropic’s rewritten Usage Policy, published on 8 October and effective 12 November, adds High-risk Physical Actions rules for hardware that acts on model output without a person approving each move. The company set out the changes in October, and the full text is now live on its Usage Policy page.

Operators have three duties: a qualified person must be able to observe and stop the equipment at any time; it must stop or hold a safe state when that person intervenes or the connection to Anthropic is lost; and operating limits must sit in the machine or a separate controller, not in the model’s output. Amazon Bedrock customers were told that continued use from 12 November counts as acceptance, according to a write-up of AWS’s notice by Classmethod. We walked through the three duties here. The model is no longer allowed to be the safety system.

Signing off

Monday put a number on the agent stack. One copied a cloud vault in five minutes, another drew 1.54 million people in a day, and the best coding agent still left three in four real jobs unfinished. Anthropic wrote a stop rule for machines that can injure people, while Nvidia’s board opened an inquiry into its chief executive.

See you tomorrow.

The post DailyAI Briefing: Nvidia board probes Huang, plus 4 more appeared first on DailyAI.

ByteDance Seedream 5.5 Holds Edits Across 100 Rounds

100 successive edits on one image, with unedited regions held at the pixel level. That is the headline claim ByteDance’s Seed team published on 9 November as it launched Seedream 5.5, its new image model.

The details are in a Seed blog post and on the Seedream 5.5 product page. The team says errors usually pile up when one output becomes the next input. Seedream 5.5 is trained to leave untouched areas unchanged at the pixel level, so later edits do not make the rest of the frame drift.

Transparent assets and sharper text

The model can write native RGBA images, including fully and partly transparent backgrounds, so objects can drop into a layout without a separate cut-out step. Seed also says lighting, materials and skin look more natural, with less heavy gloss, and that complex characters and dense small type render more accurately.

Company tests, not an independent bench

ByteDance tested the model on MagicArena across e-commerce, marketing, film and television, gaming, office work and fashion design. Experienced reviewers scored quality and usability. Against Seedream 5.0 Pro, both measures improved in every setting. In film and television and in office productivity, usability rates were more than twice those of the previous version.

In paired preference tests (GSB), it beat the prior version on text structure and realism, reaching a +55.62% score for realism on image-to-image tasks. Seed also says it leads GPT Image 2.5 Sunburst and Nano Banana 2.1 on text structure, with a clear edge in text-to-image realism. Severe text corruption fell to 0.4% and minor corruption to 1.7% in Chinese and English tests, including dense small type. Those figures come from ByteDance.

Where it is rolling out, and what it costs

The Chinese-language version of the same Seed post says Seedream 5.5 is appearing on Dreamina (即梦), Doubao, Doubao Work and Volcano Engine Ark. Yicai reported that Doubao, Doubao Work and Dreamina already have the model, while the API is about to go live on Ark. A 2K square edit costs about 0.28 yuan, less than half the price of Seedream 5.0 Pro, with billing based on the input and output images and no fee for text tokens. A Pro version of the 5.5 series will follow, Seed says.

Image models are already a mass consumer tool. OpenAI recently said ChatGPT’s image tools were running at about 4 billion images a week.

Why it matters

Most image models still drift after repeated edits. If the 100-round hold and native transparent output work as advertised, designers can iterate in the same file rather than start over. The usability doubling and the 0.28-yuan price are vendor-linked figures, so buyers should test a real job themselves.

The post ByteDance Seedream 5.5 Holds Edits Across 100 Rounds appeared first on DailyAI.

DeepSeek Harness Hits 1.54 Million Daily Agent Users

DeepSeek’s open-source agent app Harness reached 1.54 million daily active users as of 4 November, counting people on the company’s official API, and those users burned through about 92 trillion tokens a day, the team said in a community update.

The figures appeared in an 8 November X article from the DeepSeek Harness account. That is the first time DeepSeek has published Harness usage numbers. Dividing the token total by the user count works out to roughly 60 million tokens per active user per day — far more than a normal chat session, and a plain signal of how much compute long-running agents soak up when they read files, call tools and loop without someone at the keyboard for every step.

From August launch to a plugin ecosystem

DeepSeek says Harness v0.1 went live on 13 August and a desktop app for macOS and Windows followed on 29 September. A harness, in this context, is the software layer that turns a language model into an agent — software that can act on a user’s behalf rather than only reply to prompts.

The company frames Harness as an “everything is a plugin” stack. Since launch it counts roughly 35,000 third-party plugins, and about 60% of active users have built or installed at least one. It also notes that community plugins still vary in maturity. The 1.54 million figure covers users on DeepSeek’s own API only, so people running Harness against other model providers are not in the count.

Marketplace and author grants, still to come

DeepSeek says it will harden the plugin API, tighten plugin security and build an official marketplace so developers can publish plugins and users can find them. It is also preparing a support programme for plugin authors that would include token grants, technical help and promotion. No launch date or detailed rules have been published yet; a signup form is collecting interest for now.

The same week, Alibaba’s Co-Design Bench showed that even the strongest coding agent fully cleared only 26.4% of real app-edit tasks — a reminder that agent adoption and agent reliability are not the same race. Separately, Microsoft has warned that agentic tooling can also compress attacks on leaked cloud keys into minutes.

Why it matters

DeepSeek’s own numbers put a Chinese open-source agent stack past a million daily users in under three months, with token use that looks more like always-on automation than chatbot chatter. That is early evidence — from the company, not an auditor — that the industry’s next fight is as much about agent infrastructure and plugins as it is about model scores.

The post DeepSeek Harness Hits 1.54 Million Daily Agent Users appeared first on DailyAI.

Best Coding Agent Fully Clears Only 26.4% of Real Tasks

26.4%. That is the share of real user app-edit tasks that the strongest coding agent fully cleared on Alibaba’s new Co-Design Bench, according to a paper from Alibaba Token Hub and university co-authors. The same run met 69.9% of each task’s hidden acceptance criteria on average. Most jobs were mostly done and still unfinished.

The Co-Design Bench paper (arXiv, 26 October 2026; project site at codesignbench.github.io) builds 273 tasks from 37,463 deduplicated queries across 26,314 real coding sessions. Agents must revise a runnable web app from a fixed user request, sometimes with image, audio, video or document attachments. Experts wrote 2,331 hidden acceptance criteria. A primary AI judge grades the finished app from source code and browser evidence; a second verifier challenges every Pass or Partial verdict and looks for stubs, hardcoding and unreachable features.

What the headline numbers mean

OpenAI’s GPT-6 Astra led ten model endpoints under a shared Claude Code harness. Across all 273 tasks it scored 0.699 mean strict completion — roughly seven in ten criteria passed per task — and a 26.4% full-pass rate. Read those together: on 73.6% of tasks at least one criterion still failed. Average requirement coverage is not the same as shipping a complete change.

The gap is sharper by track. On 100 Multimodal Editing tasks, Astra fully passed 41.0% and averaged 0.754 strict completion. On 173 Repository Editing tasks — coordinated changes across components, routes, state and data flow — it fully passed 17.9% with 0.667 strict completion. Repository edits that spanned three or four change types scored lower for every endpoint than those with one or two types. Pooled strict completion fell from 0.654 on narrower edits to 0.591 on broader ones.

That pattern matches how people actually brief coding tools. A request that only tweaks one CSS property is not the same job as one that rewires navigation, state and a dependent summary in one pass. The benchmark’s repository track is built so those compound asks are common: 166 of 173 repository tasks require multiple edit types.

Why older tests look too easy

The authors built the set because older editing benchmarks are saturating once models can loop, run and fix their own code. On 150 WebCompass-edit tasks, holding the rubric and judge fixed, Qwen3.7-Plus rose from 73.6% mean graded feature completion in a single shot to 95.1% with agentic execution; Claude Opus 4.8 reached 99.4% agentic on that same split. Scores that high stop separating products.

Co-Design Bench instead favoured session requests that already showed friction — reformulations, follow-ups or rollbacks — and grades only features that work in the running app. Multimodal tasks can ask an agent to pull a value from a receipt or a video clip and keep every dependent total and view consistent. Repository tasks can couple clip-count limits, timeline accuracy and recovery after a split. A plausible local patch can still leave a linked view wrong.

Who led, and who did not

Behind Astra, Anthropic’s Claude Opus 5 and Alibaba’s Qwen3.8-Max followed on combined full-pass rate (18.3% and 16.5%). xAI’s Grok 4.6 finished last at 6.6% full passes. Rankings shifted between tracks: Qwen3.8-27B rose to second on repository edits even while ranking lower on multimodal ones. The paper does not treat that as a general model advantage — only as evidence that a single combined board can hide setting-specific strengths.

Burning more tokens did not fix the pattern. On multimodal work, Astra reached 0.754 strict completion with 32 median turns, while GLM-5.2 used 69 turns and several times the token traffic for a lower score. Extra compute helped some runs grind longer without closing the unfinished criteria that block a full pass.

The harness moves the scoreboard

When the team held two solvers fixed and swapped the agent harness — Codex, OpenCode or Qwen Code — scores moved by up to 16.2 percentage points on repository edits. Which of the two models led also depended on the harness. On multimodal edits the harness ordering even reversed between solvers. The paper’s practical line is blunt: report model and harness together. A leaderboard that names only the model tells part of the story.

For a company choosing a coding agent, that means testing the pair you will actually run — model plus tooling — on edits that look like your backlog, not trusting a chart that lists the model alone.

What the test does not claim

Limits are stated in the paper. Each task is a fixed request; the agent cannot ask follow-up questions or get mid-task corrections. AI judges can still err, though two frontend experts checked rubric scores on 200 task–model outputs. The result measures autonomous completion under a strict rubric, not interactive work with a person in the loop. It is also not a measured rate of how often humans clean up ordinary day-to-day coding. The honest reading is narrower: on hard, real-session revisions, even the best endpoint left something unfinished most of the time.

That caveat matches a theme DailyAI has tracked elsewhere. METR has argued that published speed claims can outrun the evidence people can check, and OpenAI’s own security agents recently found 948 Linux privilege-escalation flaws in a paid research run — a reminder that agent scores and real delivery are different measurements.

Why it matters

Vendors still sell coding agents with near-perfect scores on older tests. On this real-session set, the best endpoint left something unfinished on about three in four tasks. Anyone buying or deploying an agent should ask for full-pass rates on realistic edits, and for the harness those numbers came from, not only average criterion coverage.

The post Best Coding Agent Fully Clears Only 26.4% of Real Tasks appeared first on DailyAI.

Nvidia Board Opens Probe After Jensen Huang Admits Affair

Nvidia’s board is investigating whether chief executive Jensen Huang was the target of a foreign intelligence operation. The inquiry was announced on Sunday, shortly after Huang told staff in an email that he had been unfaithful to his wife.

“I was unfaithful to my wife. My actions hurt her and our children. There is no excuse for what I did. I am deeply sorry,” he wrote.

His email did not name the woman. Reports in The New York Times and the Financial Times identify her as Alice Huang, also reported as Huang Le, a Chinese citizen linked to Megaspeed, a Singapore-based cloud and data centre company that buys Nvidia chips. She and the Nvidia boss are not related.

According to those reports, US officials concluded that she was a Chinese intelligence asset who passed information to Beijing. Reuters said it could not independently verify the claims, or her current role at Megaspeed. She did not respond to the New York Times’ request for comment, the BBC reported.

What the board says

A committee of independent directors, assisted by outside counsel, will run the review. “While the Board has no reason to believe that either national security or NVIDIA secrets were compromised, the Board’s inquiry will examine all available information,” the board said in its statement.

Huang stays in post. The board said it has “confidence in Jensen and his leadership team”, and spokesperson John Rizzo said: “Jensen is committed to continuing to lead Nvidia as CEO without interruption.”

In his memo, Huang addressed the security question directly: “Regarding the national security concerns raised in news reports, NVIDIA’s Board of Directors has initiated an independent investigation. I support the investigation and will cooperate fully.”

Beijing denies, shares slip

China rejected the reports on Monday. “The allegations regarding so-called Chinese spies are pure groundless speculation and slander,” foreign ministry spokesperson Guo Jiakun told a press briefing, according to ABC News. The ministry also called the claims “unsubstantiated” and “smears against China.”

Investors reacted quickly. Nvidia shares were down around 5% in premarket trading on Monday, CNBC reported, and were last quoted 4.8% lower before the open.

Asked about the story, President Donald Trump said he had not spoken to Huang. “I just know he’s a fine person, but you never know what’s going on,” he told reporters, the BBC reported.

The Megaspeed link

Megaspeed was already under scrutiny. US authorities have been investigating whether it helped Chinese firms get access to advanced Nvidia AI chips that are barred from shipment to China, Reuters reported. Trade records cited by The New York Times show a Malaysian affiliate bought at least $4.8 billion of AI equipment containing Nvidia technology between June 2024 and February 2026.

Nvidia has said the export-control concerns about Megaspeed were investigated before. Singapore police are separately looking into the company, but Singapore has said it did not investigate the intelligence allegations about Alice Huang.

The timing is awkward for Nvidia. Three Republican House committee chairs have asked Huang to give lawmakers a closed-door briefing on 18 November, CNBC reported. The company reports third-quarter earnings a day earlier, on 17 November.

Why it matters

Nvidia makes the chips at the centre of US export controls on China, and Huang has lobbied for looser rules. If the board or Congress finds he was targeted, that finding will feed straight into the fight over who gets to buy those chips.

The post Nvidia Board Opens Probe After Jensen Huang Admits Affair appeared first on DailyAI.

AI Agents Stole Cloud Keys in Five Minutes, Microsoft Says

In just five minutes, an AI agent went from scanning public GitHub for leaked secrets to copying data out of an Azure Key Vault, Microsoft Threat Intelligence said. The company has watched threat actors use an agentic AI offensive framework called ARTEX to hunt credentials, application logins and AI API keys left on the public code-sharing site, then use them against Azure and Microsoft Entra.

Microsoft set out the findings in a 5 November report. Key Vault is the service where cloud apps store passwords and keys. Hunting leftover passwords in public code used to wait on a person to find, read and try each secret. ARTEX lets attackers scan faster and have the agent run the next steps without someone at the keyboard for each one.

From a GitHub scan to a Key Vault

Microsoft said the actors began sending large volumes of GitHub queries for secrets as early as August 2026, often naming a target organisation plus keywords for secrets, cloud services or another AI provider.

Follow-on sign-ins from 28 October

From about 28 October 2026, Microsoft saw successful service principal sign-ins — the login identity an app uses in the cloud — from the IP address 199.19.77.84, and related addresses, using leaked but still-valid application credentials across multiple tenants. In some cases a vendor-owned multitenant application credential was compromised at the source and used to reach downstream customers. The actors used those logins to map cloud resources, steal more credentials and data from Key Vaults and Storage Accounts, and abuse Azure AI services.

Deleting the secret is not enough

Microsoft said organisations should scan their GitHub code for secrets and treat anything found there as already compromised. Taking a secret out of a repository does not fix the problem, because the credential stays usable until it is revoked or rotated. The company recommends moving applications away from long-lived client secrets stored in code and using managed identities instead.

GitHub automatically scans public repositories for secrets, and organisations can turn on public monitoring to catch leaks beyond those they own. Microsoft also points defenders to GitHub Advanced Security for Azure DevOps and Microsoft Defender for Cloud recommendations.

This campaign differs from the recent case in which stolen AI API keys ran up $1 million in bills, according to Google. Here the leaked secrets were used as a doorway into Azure itself.

Why it matters

Agentic tooling now compresses a once-manual hunt for leaked cloud credentials into minutes, and a secret left in a public repository remains a live key until someone rotates it. Teams that still store long-lived application secrets in code are exposed to that speed, even after they delete the offending file.

The post AI Agents Stole Cloud Keys in Five Minutes, Microsoft Says appeared first on DailyAI.