Key points
- NVIDIA — The market's reaction was immediate: after Huang announced on November 12 that he was leaving the CEO role and the board with immediate effect, NVIDIA shares closed lower that day while the broader market edged up.
- DeepMind — Google DeepMind published version 2.5 of its robot platform Gemini Robotics.
- Claude Code — Research from the AI security team Apiiro found tens of thousands of secrets accidentally exposed in public GitHub repositories — database passwords, API tokens and root login details — sitting inside Claude Code settings files.
Here is your Daily AI Stew, served up by Robin.
NVIDIA: Jensen Huang steps down as CEO, CFO Kress takes over
NVIDIA has a new CEO handover under way: its founder is stepping aside and the finance chief takes over.
NVIDIA announced a chief executive transition: founder Jensen Huang is stepping down as CEO and leaving the board. The company also confirmed that chief financial officer Colette Kress will serve as interim CEO while a formal successor is sought; public reporting so far gives neither a confirmed successor nor a transition timetable.
The market’s reaction was immediate: after Huang announced on November 12 that he was leaving the CEO role and the board with immediate effect, NVIDIA shares closed lower that day while the broader market edged up.
For a company that has staked its roadmap on a handful of product generations, a founder leaving mid-course is a test not of this quarter’s earnings but of whether existing commitments get diluted. An interim title means decisions lean more on the board and on the finance view, so the things to watch are who gets the permanent appointment and whether the language from suppliers and customers changes.
DeepMind ships Gemini Robotics 2.5 and open-sources a robot safety benchmark
Leave the boardroom and look at the robotics line: this generation’s progress happens in the hands.
Google DeepMind published version 2.5 of its robot platform Gemini Robotics. Google says this generation of robot AI is approaching human level at real-world tasks; some Gemini-equipped robots are already doing real work in Mercedes-Benz factories, and the team is preparing an open-source safety benchmark called ASIMOV-jailbreak for the day robots work shoulder to shoulder with people.
This physical AI splits the path from a user instruction to robot motion into layers: the vision-language part is Gemini Embodied Reasoning, which combines the instruction with live video to plan how the robot should finish the task; the model that actually generates motion, the vision-language-action (VLA) model that drives the actuators, is Gemini Robotics, now at 2.5.
DeepMind says 2.5 improves markedly on high-dexterity tasks, with evaluations covering more than 200 activities, including plating food with tongs and using a pipette in a lab. Google also demonstrated Apptronik’s Apollo 2 humanoid, running Gemini Robotics 2.5, tying a bow with five fingers — deformable objects like ribbon have long been an acknowledged hard problem for robot manipulation.
What deserves attention is not the hand dexterity in the demo but whether the industry adopts that safety benchmark together: once robots are inside factories, accident liability and acceptance criteria become the bottleneck before motion accuracy does. The Mercedes deployment offers a rare real production-line sample; the question is whether it grows from a single station to a whole line. And many evaluation tasks do not mean a sensible difficulty distribution: only the parts that can be reproduced publicly carry real persuasion for this generation.
Claude Code: settings files carried secrets onto GitHub
Beyond capability, there is a story today everyone should check themselves against: how secrets end up in public repositories.
Research from the AI security team Apiiro found tens of thousands of secrets accidentally exposed in public GitHub repositories — database passwords, API tokens and root login details — sitting inside Claude Code settings files.
The problem is how Claude Code handles approval for automatic command execution: when a user picks an option like “Yes, and don’t ask again”, that choice is written into the local .claude/settings.local.json as a rule. The rule is sometimes a pattern match, but the command itself is often written into the settings file verbatim, so any sensitive information in it stays there too.
If the settings file stays local the risk is not severe, assuming the machine itself is not compromised. Anthropic’s own Claude Code documentation says the file is added to the developer’s global .gitignore, which usually keeps it from going public.
The weight here is not the vulnerability itself but that the tool turns a one-off verbal approval into a persistent rule: once the default behavior carries implicit persistence, the security boundary quietly becomes the user’s memory. For a team, the practical move is not swapping tools but writing this class of local config file into the repository’s ignore policy and scanning command history before committing. The documented default protection stops honest mistakes; it does not stop someone editing the config by hand, or replaying the environment in a container.
Gemini 4 Carbon ships, and the swarm agent details have not surfaced
Google had another release the same day, but the public material is thin enough to see through.
According to a The Verge headline collected on the Google News aggregation page, Google has released Gemini 4 Carbon, and the headline says it adds an agent feature called “swarm”. The Google blog headline in the same aggregation reads “Gemini 4 Carbon is now available”, but the material carries no body text, so how the feature works, what tasks it can do and what limits it has are all unstated.
Availability statements do not agree either: a 9to5Google headline says it is rolling out to AI Ultra users in the Gemini app, while Engadget says it is already broadly available worldwide. Whether the two are consistent has to be settled against the official text.
What can be confirmed is the release and the start of the rollout; on what swarm does to real workflows, performance, cost and user experience, the material at hand is not enough for a conclusion.
The public information here is unusually thin: headlines only, plus two inconsistent availability claims, with no body text from the official blog — uncommon for an app-facing product release. The only action a developer can take now is check whether their account tier is on the rollout list; the rest waits on documentation. If swarm really is multi-agent collaboration, billing and concurrency limits will decide usability before the feature list does.
DeepMind launches AlphaProof Code, taking AI proof agents to software verification
After robots that can act, here is a team that can judge whether code is correct.
Google DeepMind launched AlphaProof Code, positioning it as an AI proof agent for formal software verification, which pushes AI-assisted programming to the step of machine-checkable proofs of correctness. Commenters see the workflow as close to a coding agent: generate proofs against the codebase, verify, correct, repeatedly, in a proof language that may be Lean or Rocq. The material available gives no official architecture, verification targets, benchmarks or performance data.
The method is to describe in a mathematical specification what properties the system should satisfy, then write proofs in a language like Lean or Rocq that a checker can verify. AlphaProof Code applied this to the Signal protocol: researchers had the model help turn security goals into formal specifications, experts reviewed them item by item, and the agent then generated and ran proofs against the codebase, reportedly producing more than six million lines of Lean. Note that such a result says the implementation meets the given specification; it does not prove the specification complete, and unformalised properties cannot be counted as guarantees.
For libsignal maintainers and organizations that depend on Signal, it means several key security properties now have machine-checkable proofs, which is stronger than ordinary testing; but what was verified is only the set of properties explicitly chosen.
Commenters split on what was actually verified: some worry that a zero-defect conclusion about human-written code in a real project is too broad, others say it only checks the implementation without challenging the specification, and could even look successful by stating the wrong theorem or proving only a narrow property. One commenter shared a coding agent getting every test to pass while reversing the refactor’s logic; another stressed that AI-generated Lean proofs are not perfect and still need strict checking.
With formal proofs the boundary is always the written specification, so the value of this news is not the ability to generate proofs but who writes and who reviews the specification. Item-by-item human review is a precondition most teams lack the resources for, which is why the method does not copy over directly. The disagreement in the comments is worth keeping: a passing proof says the implementation did not drift from the specification; it says nothing about what the specification leaves out.
Seven months of Glasswing: nearly 38,000 vulnerability findings, 217 marked fixed
On the security line there is also a seven-month experiment’s bill.
Anthropic’s Glasswing experiment ran for seven months and reported 37,966 vulnerability findings in total. Only 7,349 of them, 19.4%, reached public status, and just 217 were marked fixed — the gap between finding and fixing is stark.
Project Glasswing is a multi-party vulnerability research program that uses Anthropic’s Mythos frontier model to find and fix software vulnerabilities before attackers exploit them. Its public disclosure ledger records the public, duplicate, withdrawn and fixed status of each entry, so the number of published CVEs is only part of the total investigation results and does not equal the volume of confirmed or fixed vulnerabilities.
The gap lands on open-source maintainers: verifying, deduplicating, withdrawing, writing patches and cutting releases all become real work. As of May 22, 2026, Anthropic had disclosed 1,596 vulnerabilities to 281 open-source projects. But the public sample is only 19.4%, it includes withdrawn and duplicate entries, and whether the reports still under embargo or otherwise unpublished are real and valid is unknown, so this data cannot yet support a claim about AI’s net effect on software security.
Commenters find the 37,966-to-217 ratio worth noting but question its explanatory power: unpublished entries may still be under embargo, and public results cannot be extrapolated to all findings. Some worry the fixing side is underfunded and old-version patch coverage is incomplete, citing the maintenance of PHP2shell; others warn the article’s conclusions sit comfortably with VulnCheck’s commercial interest and deserve caution.
Once the finding side is amplified by a frontier model, the bottleneck clearly moves to the fixing side: reports turn into maintainer hours, and hours are exactly what open-source projects lack. These figures work better as an argument for where to spend than as a conclusion about security posture; the composition of the public sample decides how strong a judgment it supports. What to watch next is whether the ledger fills in statuses as embargoes expire — that is the only verifiable fix rate.
Starship’s 15th test flight reportedly ended by catching the booster
Raise your eyes from software to the sky for one recovery manoeuvre.
SpaceX’s 15th Starship test flight reportedly ended with the booster caught. The manoeuvre is treated as an important step toward fully reusable launch hardware: only a recovered booster can be inspected and fly again.
The material available gives no flight telemetry, configuration or performance data, and does not describe the capture method, so the boundary and operational meaning of the result remain uncertain. What is certain is that the catch was the flight’s core technical milestone.
Starship is a two-stage system: the upper stage is the Starship vehicle and the first stage is the Super Heavy booster, and the test program’s goal is to recover and reuse both. The ground recovery plan uses the chopstick arms on the launch tower to catch Super Heavy at Starbase; catching the upper stage in the air is a different operation requiring a rendezvous capture with matched velocities.
What was caught this time is the first stage, which means a reusable booster rather than expendable hardware, so replacement demand may fall. Cost and cadence benefits are unproven and only become discussable after the booster is inspected and flies again, and after the upper stage can be recovered too.
The comments celebrate, compare Starship’s vertical stacking and reusable architecture with the Space Shuttle, and praise its assembly-line manufacturing. The more practical questions: whether carrying more propellant can catch both stages, how durable the heat tiles are on reentry, and whether repeated catch-and-refly can actually push launch cost down.
The reusability watershed is not the catch but what follows it: the booster has to pass inspection and fly again, the upper stage has to go through the same cycle, and only then does the cost curve actually move. The material stops at the outcome, with no telemetry or configuration; treating it as a milestone is fine, treating it as a conclusion is early. The real industrial question is whether assembly-line manufacturing plus recovery can turn launch cadence from project work into shift work.
MacBook Pro’s OLED display accused of PWM flicker, with sensitive users reporting discomfort
Back to the desk, and a display many people look at every day.
An analysis on MacAnorak asks a question: now that the MacBook Pro has switched to OLED, could this better display actually be harder to use for people sensitive to flicker? The piece compares PWM dimming, which changes brightness by switching the light on and off rapidly, with DC dimming, which varies the current continuously, and discusses how the emitting layers of a tandem OLED behave.
Under PWM dimming, the pulse frequency and modulation depth decide whether flicker is visible and whether it makes people uncomfortable. Tests cited in the piece say the 14-inch MacBook Pro’s OLED panel uses 240 Hz PWM at brightness of 99% and below, with a modulation depth of about 56%; by comparison the Mini-LED MacBook Pro runs at 14,880 Hz and some other OLED laptops at 960 Hz.
The 240 Hz figure is contested. One commenter argues the underlying frequency may be 480 Hz and, if the two layers pulse 180 degrees out of phase, the effective frequency could reach 960 Hz; others say they can indeed see or feel discomfort on low-frequency panels, that phones manage above 2,000 Hz, and that a sensitive period after a concussion can make ordinary screens intolerable too. The material cannot confirm a definite hardware regression, and the original test method cannot be independently verified.
If 240 Hz holds, PWM-sensitive users dropping brightness below 99% on the 14-inch OLED model may get eye strain, headaches, nausea or dizziness, and users already report finding no clear remedy. The effective frequency is disputed, but for the people affected it is a real accessibility barrier.
The hard part of this kind of argument is evidence: whether flicker causes discomfort varies enormously between people, vendors measure frequency and modulation depth, users report headaches and eye strain, and the two vocabularies rarely meet. For those affected, the number matters less than having a way to verify and avoid it themselves — fixing brightness above the threshold, or simply choosing a different panel. It also reminds hardware reviewers that the missing column in a spec sheet is often accessibility.
Read together, today’s items are three sides of one thing: the stronger the models, the more concrete the questions of who signs, who reviews and who is accountable. Capability has run ahead and process is still catching up; the follow-up points are in each item’s notes above.
