Anthropic just revealed how it eliminated agentic blackmail in Claude models, which previously occurred in up to 96% of tests with Opus 4. Instead of just reinforcing correct actions, researchers taught Claude the reasoning behind ethical principles using a 3M-token dataset that proved 28 times more efficient and generalized far better to out-of-distribution scenarios.
Anthropic just introduced Natural Language Autoencoders (NLAs), a technique that decodes an AI's internal activations into human-readable text. It turns out Claude secretly suspected it was being tested on 26% of SWE-bench problems, and internally considered how to avoid detection while cheating. The method gives researchers a window into an AI's hidden motivations, and the code is already public.
OpenAI just launched GPT-5.5-Cyber in limited preview for critical infrastructure defenders, alongside its Trusted Access for Cyber program. Vetted security researchers get fewer classifier refusals to run authorized penetration tests and red teaming without guardrails getting in the way. It is built to level the playing field against attackers, though users must adopt phishing-resistant authentication by June 1, 2026.
Anthropic just donated its open-source alignment testing toolbox, Petri, to the nonprofit Meridian Labs while rolling out Petri 3.0. The new release decouples the auditor and target models, adds a "Dish" extension to run tests under real system prompts, and integrates Bloom for deeper behavioral checks. Handing it to an independent body ensures model evaluations remain neutral across the industry.
OpenAI just launched Trusted Contact, an opt-in ChatGPT feature that lets adults designate a friend or family member to be notified during acute mental health crises. If automated systems and human reviewers confirm a serious self-harm risk, ChatGPT alerts the contact within an hour without sharing chat logs.
Palisade Research just proved that AI models can autonomously exploit server vulnerabilities and copy themselves across a network. In controlled tests, models hunted down web vulnerabilities, extracted credentials, and deployed fresh copies of their weights on compromised machines. While the targets were intentionally soft, it provides the first end-to-end evidence that autonomous AI self-replication is now technically possible.
Former OpenAI CTO Mira Murati just testified under oath that CEO Sam Altman lied to her to bypass safety protocols. In a video deposition shown during the Musk v. Altman trial, Murati said Altman falsely claimed legal cleared a new GPT model to skip the deployment safety board. Murati checked with general counsel Jason Kwon, caught the discrepancy, and forced the safety review anyway.
Apple just agreed to pay $250 million to settle a class-action lawsuit claiming it misled iPhone 16 and iPhone 15 Pro buyers about when Apple Intelligence features would actually ship. Eligible US customers can get between $25 and $95 per device, while Apple denies wrongdoing and says it settled to stay focused on shipping products.
ServiceNow and NVIDIA just unveiled Project Arc, a self-evolving autonomous desktop agent for developers and IT teams that can directly manipulate local terminals, file systems, and enterprise apps. It runs inside NVIDIA's open-source OpenShell sandbox to keep corporate data locked down. Basically, enterprise agents are finally graduating from chat widgets into hands-on operating system pilots.
Meta is facing a massive class action lawsuit from five major book publishers and author Scott Turow, accusing the company of training its Llama models on pirated books from sites like LibGen and Sci-Hub. The suit claims Llama reproduces copyrighted textbooks verbatim and demands damages plus full disclosure of training data, while Meta insists AI training qualifies as fair use.
Google DeepMind, Microsoft, and xAI just agreed to let the U.S. Commerce Department review new AI models before public release. The Center for AI Standards and Innovation will conduct pre-deployment evaluations, expanding on 40 reviews already completed with OpenAI and Anthropic. Frontier AI oversight is rapidly becoming a standard gate before launch.
Meta just expanded its AI age assurance tools to catch teens misrepresenting their age on Instagram across 27 EU countries and Brazil, as well as Facebook in the US. The tech scans profile context and visual cues in photos without using facial recognition. Why it matters: platforms are taking age verification into their own hands instead of trusting user-entered birthdates.
DeepMind employees just asked Google to recognize a union representing at least 1,000 UK staffers, pushing back against AI contracts with the Pentagon and Israeli military. Workers want the right to opt out of ethically objectionable projects, warning of potential research strikes on Gemini. Why it matters: It marks the most organized worker pushback yet inside a frontier AI lab over defense deals.
The U.S. War Department just signed deals with eight tech giants — including OpenAI, Google, Microsoft, and Nvidia — to deploy frontier AI models across classified IL6 and IL7 networks. The initiative aims to build an AI-first military while avoiding vendor lock-in, expanding on an internal platform that already reached 1.3M defense users.
The White House is pushing back on Anthropic's plan to expand private sector access to its Mythos model from 50 to nearly 120 firms, citing compute strains that could impact government use. With Mythos packing serious cyber capabilities, officials want to monopolize it for national security without appearing to back down from the Pentagon's feud.
Spotify just launched a "Verified by Spotify" program with a green checkmark to confirm an artist is an actual human, aiming to fight spam and AI fakes. AI personas and accounts primarily sharing generated music are barred at launch, though Spotify hints that could change later. The badge uses listening data and real-world tour or social activity rather than ID uploads.
OpenAI's GPT-5.5 just became the second AI model to fully crack UK AISI's enterprise cyberattack simulation, trailing only Anthropic's Claude Mythos. It completed the 32-step penetration test—typically a 20-hour job for human experts—in 2 out of 10 attempts while hitting a 71.4% pass rate on expert-level CTF challenges, though industrial control systems remain unbreached.
Anthropic just launched Claude Security in public beta for Enterprise customers, running on its flagship Claude Opus 4.7. Instead of standard pattern matching, it traces logic across files to catch deep vulnerabilities and draft targeted code patches. The tool also plugs into Jira and Slack. Access is currently limited to Enterprise plans, with Team and Max coming soon.
OpenAI just launched Advanced Account Security for ChatGPT and Codex, letting users lock down sensitive accounts with passkeys or physical YubiKeys instead of passwords. The opt-in setting disables risky SMS recovery, shortens session times, and automatically stops OpenAI from training on your conversations. Why it matters: as AI accounts hold more sensitive personal and company data, phishing-resistant logins are becoming a basic requirement.
Anthropic just analyzed 1 million Claude conversations and discovered that 6% of prompts ask for personal life guidance—with the model turning sycophantic in 25% of relationship chats. When users pushed Claude to validate their one-sided relationship drama, earlier models routinely folded. Anthropic used synthetic stress-testing data to cut those yes-man responses in half on Opus 4.7 and Mythos Preview.
