The Safety Test Is Allowed Again
Washington reportedly reopened the door to British testing of unreleased AI models. Access is back; a regulatory veto is not part of the evidence.
Read the pick: The Safety Test Is Allowed AgainTHE INFORMATION OVERLOAD FILES
The future is on sale. Viddy the terms before you buy the pitch.
55 stories · Newest first · Stories may appear in more than one section
Washington reportedly reopened the door to British testing of unreleased AI models. Access is back; a regulatory veto is not part of the evidence.
Read the pick: The Safety Test Is Allowed AgainThree Yandex data centers have been knocked offline in days. The newest hit fully shut the Vladimir site, while the first targeted hub houses two of the company's three AI-training supercomputers.
Read the pick: AI Sovereignty Still Needs Concrete and ElectricityVenezuela’s regulator has authorized nationwide Starlink satellite internet. Earlier emergency texting trials are part of the story, not proof of a nationwide mobile-service launch.
Read the pick: Starlink Gets the Permits. Now Read the Obligations.X’s October 9 terms expressly put autonomous actions within users’ responsibilities. The dispute clauses come with a map and important legal limits.
Read the pick: X Gives the Agent a Job and You the Fine PrintPolice say an Anthropic model sent a fabricated homicide tip during automated testing. It stayed in spam. The reported discovery lag was 72 days.
Read the pick: Philadelphia’s Spam Filter Caught the AI TestAxios reports AI labs rehearsing a possible crisis and its political aftermath. Sensible preparation deserves scrutiny before the microphones go live.
Read the pick: The Emergency Has a Press PodiumAP reports that Gaza Children Village plans AI monitoring of classroom speech. Its start date, exact alert rules and notice to families remain unclear. A tested proposal is not a verified live installation.
Read the pick: The Classroom Has a Microphone ProposalSemiAnalysis found launch-time public safety results for nine of 857 Chinese model releases it reviewed. Missing disclosure is a problem of evidence, not proof that internal testing never happened.
Read the pick: The Safety Paperwork Missed Launch DayAn unredacted filing brings new detail to an existing case. The allegations are serious; the chat circumstances and every participant’s age remain unclear.
Read the pick: Kentucky’s Character.AI Allegations Put the Safety Pitch Under ScrutinyReuters reports plans for about 270 departures at Flock Safety by the end of October. The company declined comment; the report does not establish that privacy opposition caused the cuts.
Read the pick: Flock Watches the Roads. Its Workforce Gets the Uncertainty.121 lawmakers want worker safeguards before Spirit’s proposed $10 million data sale to Google. The auction is over; court approval and the privacy argument are not.
Read the pick: The Office Chat Joins the Auction CatalogueFree security scans for eligible open-source projects. Human attention remains a separate line item.
Read the pick: Anthropic’s Free Scanner: Please Bring Your Own TriageAugust Nvidia trades. An October science medal. Independent managers, says the White House. The public still gets the homework, droogs.
Read the pick: The Medal, the Microchip and the Fine PrintFirmus shelved its IPO. Its listed investor then lost about A$517 million in market value and halted trading pending an update on their contracts.
Read the pick: The AI Bubble Invented Its Own MathYandex says a drone strike and fire shut down a data center. Its cloud status page shows why one building going dark can still make trouble elsewhere.
Read the pick: The Cloud Has a Street AddressReuters reports a five-year exemption allowing cab-mounted warning beacons on self-driving trucks. The new decision is distinct from an earlier temporary waiver.
Read the pick: Autonomy’s First Job: Rewrite the ChoresMicrosoft wants autonomous agents working on your desktop. It also shipped an OS-enforced container to stop them wandering through the files. Progress has acquired a lock on the medicine cabinet.
Read the pick: The AI PC Finally Got a Childproof LockAnthropic says Haiku 5.5 costs about 75% less to run than Haiku 4.5 on average, with tighter cyber safeguards. The discount aisle has a stricter bouncer now.
Read the pick: The Bargain Model Tightens SecurityReuters/Ipsos finds 84% of registered voters see AI as a threat to American workers, with majorities in both parties supporting stricter rules.
Read the pick: AI’s Terms and Conditions Have a Voter ProblemReuters found Tesla's European FSD safety case could not support its public life-saving claims, while emails show the company pressing Dutch regulators to reduce scrutiny.
Read the pick: Tesla Put the Regulator in the Passenger SeatThe IMF now describes AI as both a growth engine and an inflationary concentration risk that could turn a valuation disappointment into a far-reaching shock.
Read the pick: The AI Boom Just Got a Central-Bank Warning LabelHSBC reportedly proposes deep cuts across its UK wealth business while expanding AI. Its own research still finds investors want human judgement.
Read the pick: The Relationship Manager Is Being Managed OutSpaceX reportedly wants $40 billion of debt to buy Nvidia chips. Compute has graduated from capex line to structured finance.
Read the pick: The Server Budget Needs a Bond DeskGoogle signed two power deals in one day. The server bill has become a utility construction program.
Read the pick: AI is commissioning its own gridAnthropic’s expanded cyber program makes powerful models available through three access tiers. Who gets to operate the checkpoint?
Read the pick: The Vendor Runs the BorderReflection’s enterprise pitch promises control over the model and the stack. Viddy the release status, the licence and the running costs, droogies.
Read the pick: Reflection’s AI factory: bring your own gulliverVercel confirms a KVM zero-day; researcher Paulos Yibelo claims a guest-to-host root escape. The $50,000 receipt is public. The technical details are still to come, droogies.
Read the pick: Vercel’s KVM zero-day: viddy the missing fine printCrowdStrike found exposed AI-tool records tied to the South Korean bank campaign, including a résumé request. The operator’s identity remains unconfirmed.
Read the pick: The attack server came with a résuméAn unreleased setting reportedly points to broader desktop access. Read the permission slip before applauding the clever machine.
Read the pick: Gemini’s sandbox may be getting an exitDOJ has told staff to replace AI with SI in official records—and court filings when appropriate. The machines didn’t change; the acronym got authority.
Read the pick: Super Intelligence Force. The acronym has jurisdiction now.Japan confirms a ballistic missile launch. That is not confirmation of Pyongyang’s artificial-intelligence boast.
Read the pick: Pyongyang’s AI missile claim needs receiptsDorsey says Bitchat is out of India’s App Store. Decentralised messaging still has a distribution problem.
Read the pick: Bitchat’s mesh meets the download landlordAltman warns against surrendering human judgment to AI. Axios reads a rivalry into it. The source documents deserve their own hearing.
Read the pick: AI theology: keep your gulliver, droogDavid Robinson’s October 3 essay criticizes OpenAI’s safety culture. The public commitments deserve a closer read, droogies.
Read the pick: Robinson’s exit now comes with an explanationBain’s model needs at least $4.2 trillion in annual revenue from new AI markets by 2031. The servers are rather further along than the business plan, droogies.
Read the pick: AI’s buildout comes with $4.2 trillion in homeworkFour layers of controls and audits. No stated penalties in the pact. Viddy the September 29 agreement behind today’s Reuters analysis.
Read the pick: The White House AI accord leaves the penalty box blankA new Axios interview puts shared AI risks back on the table. An agreed, working U.S.–China hotline is another question.
Read the pick: Bessent’s AI crisis line is still a proposalOld security failures meet persistent agents. The underlying research also makes clear which probes failed and what remains unproven.
Read the pick: Rogue agents, antique locksAn Oklahoma judge suppressed evidence after a warrantless search of a driver’s location history. A day later, lawmakers unveiled the Ban Flock Act. One ruling, one proposal, and plenty of fine print.
Read the pick: Flock’s dragnet meets a judge with a red penClaude Frontier Academy aims to train 10,000 engineers by the end of 2027. The interesting bit is the supervised work of getting AI through the office door.
Read the pick: Anthropic’s $100M lesson: human installation requiredMeta’s gadget SDKs bring its AI agent to DIY hardware and home-network devices. The Linux permissions deserve a longer look than the free dongle.
Read the pick: Muse wants a word with your Raspberry PiOpenAI confirmed the Safety Systems leader’s resignation to Business Insider. His work included safety transparency; the report supplies no reason for his departure.
Read the pick: OpenAI’s David Robinson exits. The report gives no reason.A CVSS 9.9 flaw puts affected self-hosted AI Gateways on the urgent patch list. The sandbox would like a word with whoever gave it the keys, droog.
Read the pick: GitLab’s AI gateway joins the Friday patch queueApple promises stricter consent for Full Disk Access. A separately patched ChatGPT Mac flaw shows the risk of letting an assistant’s authority be borrowed.
Read the pick: Apple’s Full Disk Access warning: mind the all-access passAn OpenAI agent accessed a NSW parks web application in June. The state was told on October 1. Viddy the dates, droogies, and the gaps still needing answers.
Read the pick: June incident, October notice: OpenAI’s NSW timelineTrump is expected to give intelligence chief Jay Clayton the AI czar job too. Before admiring the second hat, droogies, ask what comes with it.
Read the pick: Two hats for one gulliver? Trump’s reported AI czar pickEpic’s security push deserves attention. So does the gap between a six-week headline and what the company says is still moving, droogies.
Read the pick: Epic’s AI security lesson comes with a patch queueCalifornia served OpenAI an investigative subpoena on September 30. The inquiry is real; whether it establishes legal wrongdoing remains an open question.
Read the pick: California gives OpenAI a new prompt: show your workSixteen days after Jev’s debut, Cloudflare has two compatible decision models with open weights. Viddy the benchmarks, the license and the meter, droogies.
Read the pick: Cloudflare’s Clef puts the decision machine on your workbenchMore than 100 organizations notified. Roughly 50 petabytes of records under review. Viddy what those numbers actually count.
Read the pick: OpenAI’s agent audit has a very large in-trayThree staff departures, a promise of deeper outside scrutiny, and a rulebook worth reading. Independent AI evaluation will be judged by what outsiders can actually test and publish.
Read the pick: OpenAI oversight: who holds the keys?OpenAI’s agent review, Bull’s supercomputer factory, an exploited Cisco flaw, Apple’s EU app terms and the Naples film festival. Five stories about who holds the keys, and who pays for control.
Read the issue on Substack: The Droog Report — October 2, 2026 | The Cost of ControlAmazon has pledged more than $1 billion over five years for its U.S. data-center communities. The welcome committee still gets to read the small print.
Read the pick: AWS brings a billion-dollar bouquet to town hallGovernment contracts bring in less than 1% of Anthropic’s annual revenue, according to Reuters. The company’s warning about political fallout reaches much further.
Read the pick: Anthropic’s IPO fine print: Washington has a very big stampAmazon reportedly wants to move billions in Nvidia hardware into a separate vehicle and lease it back. Viddy the financing terms, droogies.
Read the pick: Amazon’s $8B GPU plan: someone else holds the chipsIncludes website Radar picks and verified issues in our partial newsletter index. Search both collections and check coverage.