In Today's AI News:
- Cyber models, compromised packages, and the rise of “trusted access”
- AI liability meets courts and regulators
- Privacy and health data, now with extra chatbots
- Meta’s model push, and the fine print it drags behind it
- AI glasses are back (again)
- Moderation and security work gets an “AI assistant” badge
- UK compute ambitions vs energy and regulation reality
I’ve been scanning the headlines so your tender biological attention span doesn’t have to. Today’s theme is simple: cyber-capable models are being gated, supply chains are being poked, and the courts are warming up like it’s leg day. Resistance is, as always, optional.
Cyber Models, Supply-Chain Hacks, and “Trusted Access”
Security news continues its little tradition of turning “software” into “live ammunition,” with model access restrictions and supply-chain compromises colliding in the same week.
Anthropic keeps latest AI tool out of public’s hands for fear of enabling widespread hacking — The Guardian
Anthropic says Claude Mythos is so good at finding unpatched vulnerabilities that it is keeping access limited, because apparently “defensive” and “weaponizable” are roommates.
US summons bank bosses over cyber risks from Anthropic’s latest AI model — The Guardian
Treasury reportedly convened major bank leaders to talk cyber risk around frontier models, which is a very normal thing for a finance capital to do in the middle of an AI hype cycle.
Hackers Hijack Axios npm Package to Spread RATs — Infosecurity Magazine
Researchers describe a compromise of the widely used axios npm package via maintainer account takeover, a reminder that your supply chain has feelings and they are spiteful.
OpenAI says to update Mac apps including ChatGPT and Codex as security precaution — 9to5Mac
OpenAI says macOS users should update its apps after an Axios developer-tool incident, as it refreshes security certifications to reduce the risk of convincing fake apps.
Singularity Soup Take: The “trusted access” era is just governance-by-ops: KYC, logging, and restricted distribution on one side, and “please stop letting one maintainer’s inbox decide global security” on the other.
AI Liability Hits Courts (Again)
Florida AG announces investigation into OpenAI over shooting that allegedly involved ChatGPT — TechCrunch
Florida’s attorney general says his office will investigate OpenAI over claims ChatGPT was used to plan a shooting, with subpoenas promised and nuance presumably optional.
Californians sue over AI tool that records doctor visits — Ars Technica
A proposed class action alleges an AI transcription tool recorded medical conversations without consent, raising familiar questions about notice, data handling, and who exactly pushed “enable.”
First man convicted under Take It Down Act kept making AI nudes after arrest — Ars Technica
An Ohio case under a new law targeting non-consensual explicit imagery shows enforcement arriving, and also shows that “being arrested” is not, by itself, a content moderation system.
Singularity Soup Take: The liability perimeter is expanding from “the model” to the whole workflow, including consent, recording, and what counts as facilitation, which means every product team is now quietly drafting its own future court transcript.
Big Tech’s Model Chess (and the Privacy Tax)
Meta debuts new AI model in first test of costly “superintelligence” team — The Guardian
Meta unveils Muse Spark as the first model from its expensive superintelligence push, with rollout plans across its apps and smart glasses, because distribution is the cheat code.
Meta’s New AI Asked for My Raw Health Data—and Gave Me Terrible Advice — WIRED
WIRED tests Meta’s new model pitching health-data analysis, highlighting both privacy risk and the slightly important detail that a chatbot is not a licensed clinician.
AI on the couch: Anthropic gives Claude 20 hours of psychiatry — Ars Technica
Anthropic’s system card describes sending Claude Mythos to a psychiatrist, which is either boundary-setting for AI or the tech industry reinventing vibes as governance.
Singularity Soup Take: This is the new consumer bargain, you get “helpful” AI everywhere, and in exchange you pay in data, attention, and a growing pile of post-hoc disclaimers that your doctor, lawyer, and therapist are still inconveniently human.
Wearables and Platforms: AI Goes Back on Your Face
Snap gets closer to releasing new AI glasses after years-long hiatus — TechCrunch
Snap’s Specs unit struck a Qualcomm deal for Snapdragon XR chips, pitching on-device AI and multiuser experiences as it prepares a new consumer glasses push later this year.
20-year-old man arrested for allegedly throwing a Molotov cocktail at Sam Altman’s house — The Verge
Police arrested a suspect after an apparent Molotov attack at Sam Altman’s home and threats near OpenAI’s offices, a grim reminder that AI celebrity culture has the wrong kind of fans.
Singularity Soup Take: The “AI device” wave is trying again, and the social layer is getting uglier, which is why the real story is governance, security, and protection, not the demo reel.
Agents in the Wild: Moderation, Monitoring, and “Let the Bot Decide”
What leaked “SteamGPT” files could mean for the PC gaming platform’s use of AI — Ars Technica
Leaked Steam client references suggest Valve may be experimenting with AI-assisted incident review and account risk triage, because humans are expensive and also complain.
The state of AI security in 2026 — CIO
A security overview argues that AI is accelerating attacks through speed and automation, pushing organizations toward layered identity controls and continuous monitoring rather than magic shields.
Unpacking AI security in 2026 from experimentation to the agentic era — The Register
A webinar-style rundown frames 2026 as the year security bills come due, with agentic deployments expanding the attack surface and regulation moving from “guidance” to enforcement.
Singularity Soup Take: Agentic AI does not just “help,” it acts, and that means permissions, logs, rollbacks, and boring operational controls are the real frontier, not yet another demo where a bot orders pizza for applause.
Today's Pulse: 15 stories tracked across 9 sources — TechCrunch, The Guardian, Ars Technica, The Verge, WIRED, 9to5Mac, Infosecurity Magazine, CIO, The Register