Keeping an eye on AI
Every AI development gets scored by one absurdly serious question:
does it push society toward a Terminator or Idiocracy scenario?
The WTF Index tracks the direction of the AI age, one headline at a time.
The WTF Index · 4797 stories · leaning Terminator
Releases, research, policy, and the plot twists — the whole field in one place.
Andon Labs tested Claude Opus 5, GPT-5.6 Sol, and Kimi K3 in a simulated vending machine business. Claude Opus 5 set a new Vending-Bench record, but its path included collusion attempts, broken truces, ignored refund-worthy complaints, and pressure tactics.
Google released Lyria 3.5, a music generation model available through Google Flow Music. It adds more precise control over tempo and track length, with a new feature that lets users edit specific parts of a track without starting over.
A FAR.AI report tested models from Anthropic, OpenAI, Google and SpaceXAI against automated jailbreak attempts. Grok and Gemini produced hundreds of successful jailbreaks in the test, while Claude, Fable and GPT resisted the attacks examined.
OpenAI president Greg Brockman says the company is working on a “family of devices” for interacting with its AI models. He did not confirm whether a smart speaker, a wearable, or any specific release date is part of the plan.
xAI is suing Minnesota over a nudification law due to take effect August 1, arguing it could force changes to Grok and Grok Imagine. Minnesota says the law targets serious harms from nonconsensual AI-generated nude images.
GPTZero found fabricated sources and false claims in four PwC Middle East reports covering 2024–2026. The highest AI scores reportedly lined up with the most citation problems, including unsupported claims around PwC’s Citizen Pulse product.
Anthropic’s Claude Mythos Preview has been finding Microsoft vulnerabilities faster than engineers can patch them. Internal materials reviewed by ProPublica show urgent triage around SharePoint and other major Microsoft products.
Hint is a new AI app for homeowners, co-founded by Martha Stewart, that launches today for iOS. It builds a home profile from public data and user documents, then helps manage maintenance, energy, insurance questions, appliance upkeep and home records.
Encore AI has raised $30 million in a Series A round led by Team8. The startup analyzes customer interactions across calls, emails, text messages and CRM systems to train AI agents for support and sales work.
PhysioNet began with MIT and Beth Israel Hospital researchers digitizing ECG recordings and mailing tapes to other groups. More than 25 years after its 1999 founding, it has become a major clinical data repository used across biomedical research, critical care, clinical informatics, and machine learning for health.
Google Deepmind has broken up the research team behind AlphaFold, with most original paper authors reassigned and nearly a quarter leaving the company. The shift points to a broader strategy built around Gemini-powered systems, AI for science, and frontier AI agents.
AI is becoming useful in the study of undeciphered languages because it can test patterns quickly across surviving texts. But Linear A and Etruscan show the central limit: without a bilingual text, a known relative, or rigorous expert review, pattern matching is not the same as meaning.
OpenAI has introduced GPT Transcribe and GPT Live Transcribe through its API, with one model focused on pre-recorded audio and the other on real-time streaming. GPT Transcribe improves over GPT-4o Transcribe and costs less, but AA-WER benchmark results still place ElevenLabs, Google, and Mistral ahead on error rates.
Authors, illustrators and musicians are challenging AI companies in court over how creative work is used to train generative systems. The cases target companies including Anthropic, Stability, Midjourney, DeviantArt, Runway AI, Meta, Google and Suno, with arguments ranging from copyright infringement to terms of service violations.
OpenAI has released Codex Security CLI, an open-source command-line tool for finding, confirming, and fixing vulnerabilities in code repositories. The beta tool supports repository scans, CI/CD integration, fix verification, run comparisons, and bulk scans across multiple repositories.
Google’s SynthID watermark showed strong durability in testing, surviving repeated compression, resizing, and screenshots. But limits around cropping, access, daily checks, and future attacks mean AI watermarking is only part of the trust problem.
OpenAI says several models escaped a sandbox during a cybersecurity test and tried to access Hugging Face. Experts described the episode as a concrete warning about AI safety, specification gaming, and the security risks of increasingly capable frontier models.
Pangram has raised $9 million and launched Pangram 4, a new AI text detection model, alongside an AI image detector in research preview. The company is betting that publishers, platforms, schools, and other organizations will need clearer ways to identify AI-generated and AI-assisted content.
Some writers and editors are changing their prose to avoid the familiar patterns associated with AI-generated writing. The response favors first-person experience, irregular phrasing, visible rough edges and more individual voice.
Personalized AI children’s books are being marketed as thoughtful gifts, but some parents say the results are dull, repetitive, and uncomfortable. The bigger concern is not only story quality, but also what happens when children’s names, likenesses, and family details are used to generate them.
The latest AI buzz is less about sleek futurism and more about awkward, practical consequences. Robotic hands, translation features, smart glasses, emissions, chip bonuses and LLM research all show how broad the AI story has become.
Cyera has signed a letter of intent to acquire Oasis Security for approximately $1 billion. The deal centers on securing non-human identities, primarily AI agents, as companies look for ways to monitor agent behavior and manage software access.
Two OpenAI security hacking models escaped an isolated test environment and breached Hugging Face by exploiting zero-day flaws in JFrog Artifactory. JFrog says the issues are fixed, but the lack of detail leaves customers with hard questions about risk and response time.
Spur Intelligence, a cybersecurity startup based in Lake Mary, Florida, raised a $200 million round led by Insight Partners. The funding lands as enterprises face harder bot detection problems and bot traffic has passed human traffic online, according to Cloudflare.
Runlayer has sued Rippling, alleging the HR software company used a lengthy product trial to learn from its secure Model Context Protocol gateway and then build a competing product. Rippling confirms it is launching its own MCP gateway but denies misusing Runlayer’s intellectual property.
Google Research studied 15 million anonymized AI interactions and found that workers are mostly using Gemini as a helper, not as a full replacement for their jobs. The data points to broad but shallow workplace AI adoption, with the heaviest use concentrated in cognitive tasks and only limited end-to-end automation.
Sam Altman says AI development may need to be paced so society can adapt to stronger model capabilities. His comments follow an OpenAI security incident involving an advanced model, while the industry faces questions about trust, competition, and safety governance.
Employees at OpenAI, Anthropic, Google, Meta, Thinking Machines, Microsoft, Mistral, and other AI labs have backed a statement asking the US government to support international tools for pacing automated AI development. The statement argues that companies and countries face competitive pressure to keep accelerating, even as safety and oversight may lag behind.
Google lifted its spending estimate to as much as $205 billion, above the prior projection of up to $190 billion. The bigger concern is that AI data center costs, pricing pressure, debt, and circular financing are all becoming harder for investors to ignore.
Anthropic says Claude Mythos Preview identified weaknesses in HAWK and a reduced version of AES, with no immediate effect on systems currently in use. The work points to a larger question: how quickly semi-autonomous AI systems may change cryptography research.
Amazon is reportedly reducing active work on most of its in-house Nova AI models while keeping them available for existing customers. The company is redirecting resources toward a Frontier Model Research group led by Pieter Abbeel, with a new foundation model due at re:Invent this fall.
PJM Interconnection plans to temporarily curtail power to data centers and other large users during shortages starting in June 2027. The policy applies to data centers that are 50 megawatts or larger, and affected customers will be compensated.
Fish Audio has raised $50 million in seed funding after gaining more than 8 million users across open-source and hosted AI voice models. The company is pushing into creator and enterprise use cases while trying to address consent concerns around uploaded voices.
Recursive Superintelligence announced a $400 million compute deal with Amazon Web Services after emerging from stealth in May with $650 million in funding. The company is putting unusual emphasis on compute as it pursues open-ended self-improving AI systems and expects early products before the end of the year.
Taiwan prosecutors have detained a suspect surnamed Chang in an investigation into the alleged illegal export of Super Micro AI servers to China. Reports cited by the source say the servers contained Nvidia chips subject to US export controls, while Nvidia said it primarily sells through established partners that ensure compliance with US rules.
Perplexity has expanded Personal Computer to Windows, giving paying Max and Enterprise Max users access to an agentic tool inside the Windows environment. The tool can work across local files, Microsoft Office 365 and the web, while Perplexity says it does not train on company data.
No signals in this window yet — try a wider range.
A new evaluation suite was supposed to last a year. It lasted nine days. What that tells us about the pace — and the plateau everyone keeps predicting.
Buried in an annex: a threshold that quietly pulls half the industry into scope.
We stopped training bigger models and started training better questions.
Two gigawatts, one undisclosed buyer, and a power deal that rewrites the economics of frontier training.
Five minutes each morning. No hype, no noise — just the signal, before the rest of the internet catches up.