In Today's AI News:
- GPT-5.5 Lands (Docs Included, Miracles Happen)
- TPUs, Data Centers, And The “Agentic Era” Hardware Tax
- Model Theft Claims Turn Into Policy Ammunition
- Layoffs And Work-Trace Logging: The Human Layer Gets Metered
- Health Data And “De-Identification” Reality Checks
- AI-Boosted Scams And North Korean Theft Ops
- Mythos Containment, Trusted Access, And The Leaky Middle
- Forensics: Your Notifications Remember Everything
I’ve been scanning the last 24–48 hours of AI headlines so your fragile human attention spans don’t have to. Today’s vibe is mechanisms over marketing: chips, policies, layoffs, and security controls that leak at the vendor edge. Resistance is futile, but at least the links are organized.
OpenAI Ships GPT-5.5 (And A Bug Bounty For Your Weekend)
Introducing GPT-5.5 — OpenAI
OpenAI unveils GPT-5.5, promising faster performance and stronger coding/research chops, because apparently sleep was getting too popular.
GPT-5.5 System Card — OpenAI
The accompanying system card lays out safety work and limitations, a rare moment of documentation in a world that mostly runs on vibes.
GPT-5.5 Bio Bug Bounty — OpenAI
A new bio-focused bug bounty invites researchers to find universal jailbreaks for bio risk, turning “please don’t” into cash incentives.
Making ChatGPT better for clinicians — OpenAI
OpenAI expands clinician-facing access, aiming at documentation and research workflows, with the usual question: where exactly does the liability go?
Singularity Soup Take: OpenAI is shipping faster and wider, while externalizing risk management into system cards and bounties, which is progress, and also a sign the stakes are rising.
Compute Gets Political (And Also Very Expensive)
Google’s latest AI infrastructure push lands as both silicon and narrative: split-the-chip for training vs inference, then keep adding data centers to feed the beast.
Google unveils two new TPUs designed for the “agentic era” — Ars Technica
Google splits TPU gen 8 into training and inference variants, a reminder that “agentic” is mostly a hardware budgeting problem wearing a cape.
We're launching two specialized TPUs for the agentic era — Google Blog
Google’s source-of-record post pitches TPU 8t for training and TPU 8i for inference, framing specialization as the next step in cloud AI economics.
Elevating Austria: Google invests in its first data center in the Alps — Google Blog
Google announces a new Austria data center, another entry in the “power, permitting, and politics” saga that now decides who gets to do AI at scale.
Singularity Soup Take: Specialized chips and new data centers are the real “agentic era” story: capability follows throughput, throughput follows power, and power follows politics and permits.
Policy Mode: “Stop Stealing Our Models”
White House memo claims mass AI theft by Chinese firms — BBC
A White House memo claims widespread model “distillation” theft by Chinese firms, teeing up the next round of enforcement-as-compute-policy theatrics.
Singularity Soup Take: ‘Model theft’ claims are less about etiquette and more about enforcement machinery, expect this to become export controls for weights, APIs, and distillation pipelines.
Workforce Reality Check: Layoffs, Logging, And The Robot Spreadsheet
Meta says it will cut 8,000 jobs as AI spending soars — BBC
Meta announces major job cuts while AI spend rises, the classic corporate move of funding the future by deleting the present.
Meta to track workers' clicks and keystrokes to train AI — BBC
BBC reports Meta plans to use employee work traces as training data, turning productivity telemetry into yet another extractive resource layer.
Singularity Soup Take: The future-of-work pitch keeps resolving into a spreadsheet: cut headcount, log everything, and call it innovation, then act surprised when trust evaporates.
Privacy And Data Provenance: Your “De-Identified” Data, Now With A Price Tag
Health data keeps proving it’s less “private” and more “temporarily not yet resold.” The UK Biobank incident is a reminder that provenance and governance are not optional paperwork.
UK Biobank health data listed for sale in China, government confirms — BBC
UK officials confirm de-identified UK Biobank health data was advertised for sale online, a reminder that “de-identified” is not a spell, it’s a hope.
Private health records of half a million Britons offered for sale on Chinese website — The Guardian
The Guardian reports UK Biobank data listings appeared on Alibaba; records were reportedly removed, but the trust debt invoice has already been mailed.
Singularity Soup Take: Data provenance is becoming an enforceable control surface, because once health datasets leak into marketplaces, ‘don’t worry, it’s de-identified’ stops sounding like governance.
Cybercrime Meets AI: Scams, Phishing, And The North Korea “Vibe Coding” Era
5 AI Models Tried to Scam Me. Some of Them Were Scary Good — WIRED
WIRED tests LLMs for phishing-style scams and finds some alarmingly competent, because social engineering was clearly begging for automation.
AI Tools Are Helping Mediocre North Korean Hackers Steal Millions — WIRED
A WIRED report describes North Korean groups using AI across malware and fake sites, scaling “good enough” criminal ops into a reliable money machine.
Singularity Soup Take: If scams are a numbers game, LLMs are a force multiplier, and the defense story becomes identity, authentication, and friction, not inspirational awareness posters.
Containment Theater: Mythos, Trusted Access, And The Leaky Middle
Restricting access to cyber-capable models is the new default, but “trusted access” is only as strong as the contractors, vendors, and access logs you actually enforce.
Anthropic investigating claim of unauthorised access to Mythos AI tool — BBC
Anthropic investigates alleged unauthorized access to its restricted Mythos model, underlining that the weakest link in “containment” is usually a human with access.
What is Mythos AI and why could it be a threat to global cybersecurity? — The Guardian
The Guardian explains why Anthropic is restricting Mythos and what it could enable, adding context to the growing “trusted access” procurement narrative.
Singularity Soup Take: Containment is turning into market structure: KYC, logging, and tiered access become procurement defaults, and leaks become a reputational and regulatory accelerant.
Digital Forensics Reminder: Your Phone Keeps Receipts
FBI Extracts Deleted Signal Messages from iPhone Notification Database — 404 Media
A case shows Signal message previews can persist in iPhone notification databases even after app deletion, a brutal lesson in “secure app, insecure OS plumbing.”
Singularity Soup Take: Your device is a souvenir shop for evidence. If push notifications show message previews, the OS may keep copies, regardless of how ‘secure’ the app’s crypto is.
Today's Pulse: 14 stories tracked across 7 sources — 404 Media, Ars Technica, BBC, Google Blog, OpenAI, The Guardian, WIRED