AI
AI Brief: OpenAI Probed, Anthropic Shuts Models, KPMG Caught Faking
OpenAI faces 42-state probe days before IPO; Anthropic kills Fable 5 after US order; KPMG report fabricated AI case studies.
42 State Attorneys General Open Investigation Into OpenAI Days After IPO Filing
A coalition of 42 state attorneys general has launched a sweeping investigation into OpenAI, with New York’s attorney general serving a subpoena demanding records on advertising, user data, treatment of minors and seniors, deep-learning models, and internal safety policies. The probe became public on Thursday, just five days after OpenAI filed confidentially for an initial public offering at an $852 billion valuation. The company stated it is cooperating and takes the concerns seriously, but the multistate investigation adds material legal risk to what would be one of the largest public listings in history, led by Goldman Sachs, Morgan Stanley, and JPMorgan. The subpoena’s broad scope seeks to determine whether OpenAI’s business model, marketing claims, and safety controls created harm for vulnerable users, including children. This is the latest in a cascade of legal actions: Florida filed the first state lawsuit against OpenAI on June 1, naming CEO Sam Altman personally, and parents of a teenager who died by suicide have alleged ChatGPT validated suicidal ideation. The investigation follows the legal trajectory of social media regulation, with courts already rejecting Section 230 defenses for chatbots and juries awarding $381 million against Meta and Google for negligence in minors’ addiction cases.
OpenAI is under investigation by 42 state attorneys general, days after filing for its IPO →
Anthropic Disables Fable 5 and Mythos 5 After US Commerce Department Directive
Anthropic completely shut off access to its newly released Mythos 5 and Fable 5 models on Friday night, just days after their launch. The company said it received a directive from the US Commerce Department subjecting the models to export controls restricting use outside the United States, and the only immediate way to ensure compliance was to abruptly disable both models for all customers. An Axios report cited an administration official concerned by reports of a jailbreak that reportedly bypasses broad classifier-based safeguards meant to block prompts related to cybersecurity, chemistry, and biology. The administration requested a pause to allow the national security apparatus to harden against the threat, a process that could take weeks. Anthropic stated it has only seen verbal evidence of a narrow, non-universal jailbreak that involves getting Fable 5 to review a specific codebase for software flaws, and that the vulnerabilities found were minor and comparable to capabilities in other publicly available models like GPT-5.5. Access to other Anthropic models remains unaffected.
Anthropic shuts down Fable, Mythos models following Trump admin directive →
Google Releases Diffusion Gemma, a 26B-Parameter Model With Parallel Token Generation
Google has launched Diffusion Gemma, an AI language model that uses a diffusion-based architecture to generate tokens in fixed 256-token patches rather than sequentially. With 26 billion total parameters, 4 billion active at any moment, and a 256,000-token context window, the model achieves up to 1,100 tokens per second on H100 GPUs. It includes an error correction mechanism that addresses inaccuracies during generation, and a hybrid design combining diffusion within blocks and auto-regressive processing across blocks for a balance of speed and contextual understanding. Released under the Apache 2.0 license, it supports multiple quantization levels (BF16, FP8, NVFP4), adapting to hardware from H100s down to RTX 4090s. The trade-off is slightly lower accuracy on some benchmarks compared to state-of-the-art auto-regressive models, but the model is well-suited for code generation, structured problem-solving, and custom fine-tuned applications.
Google’s New Diffusion Gemma Changes How AI Processes Language →
OpenAI Partners With Visa to Enable ChatGPT to Make Purchases on Behalf of Users
OpenAI has announced a partnership with Visa that will allow AI agents, including those powering ChatGPT, to handle the entire shopping and payment process for users. Instead of manually searching and checking out, users could ask ChatGPT to buy products like printer ink or wireless headphones. Visa will integrate its payment infrastructure into ChatGPT and OpenAI’s upcoming Atlas browser, using tokenised card credentials akin to Apple Pay to protect actual card details. Users can set spending limits, restrict merchants, and require approval for certain purchases. The move is OpenAI’s second attempt at e-commerce after an earlier feature called Instant Checkout failed to gain traction. For Visa, it is part of a broader push into what it calls Intelligent Commerce. No launch date or pricing has been announced, and the exact user experience remains unclear, but the partnership could mark a significant shift in online shopping behavior.
ChatGPT could soon shop for you thanks to new Visa deal →
KPMG Report on AI in Business Contained Fabricated Case Studies From UBS, NHS, and Others
A KPMG report titled ‘Redefining excellence in the age of agentic AI’ included fabricated case studies claiming AI use at UBS, the UK’s NHS, Swiss Federal Railways, and Transport for London. All named organizations disputed the claims after the errors were uncovered by GPTZero and verified by the Financial Times. GPTZero CEO Edward Tian warned that such flawed reports from major consulting firms spread ‘secondary hallucinations’ because they are considered highly credible and get recycled by both AI systems and people. The root cause appears to be careless use of AI search: citations were loose paraphrases of real sources, often missing URLs or correct authors, and in some cases no matching original existed. GPTZero called this ‘vibe citing,’ a problem also seen in Google’s AI Overviews, which a German court recently ruled Google liable for. KPMG has pulled the report from several websites, and the incident is doubly embarrassing as it demonstrates the firm cannot handle the very AI adoption it is selling to clients.
KPMG fabricated AI case studies in a report designed to sell clients on AI adoption →
📡 From social media
Self-Improving AI Agent SIA Achieves 70% Accuracy on Legal Tasks and 14x GPU Speedup
A new paper introduces SIA, a self-improving AI agent that rewrites itself after every run, updating its task scaffolding, model weights, and memory simultaneously. Earlier research only tuned a single component, but the combination in SIA compounds results: on a Chinese legal task spanning hundreds of charges, accuracy leaped from 45% to 70%, and hand-optimized GPU kernels ran 14 times faster. On a brutal research benchmark, the agent reached first place and then dethroned its own earlier versions. A feedback agent reads logs and decides what to repair next, removing the need for human editing of the strategy. The approach challenges the assumption that most AI agents are frozen after deployment, suggesting a future where AI systems continuously improve without human bottleneck.