The White House is blocking Anthropic from expanding private-sector access to its Mythos model from roughly 50 to 120 firms, citing concerns that compute strain could hurt government access. Officials want the cyber-heavy model kept close for national security, creating an awkward split between wanting Mythos's capabilities and continuing the Pentagon's hard line.
OpenAI just explained why its models started obsessing over goblin and gremlin metaphors starting with GPT-5.1. A post-mortem traced the verbal tic to reward training for ChatGPT's "Nerdy" personality, which favored creature words and accidentally leaked into the broader model via RL feedback loops. OpenAI retired the persona in March and cleaned its training data.
OpenAI just published a Cybersecurity Action Plan to help defenders keep pace with AI-powered cyberattacks. Built around five pillars—including democratizing defensive tools and securing frontier models—the initiative aims to fortify digital infrastructure for the US and allies. Why it matters: with malicious actors already scaling up automated attacks, giving defenders equal access to AI tools is becoming a matter of basic survival.
OpenAI just pulled back the curtain on how ChatGPT catches violent intent before it turns into real-world harm. The system pairs classifiers and reasoning models with human reviewers, immediately banning violative accounts and alerting law enforcement if an imminent threat is detected. It is also rolling out a trusted contact feature for adult users.
Google just signed an agreement allowing the Pentagon to deploy its AI models on classified networks. The defense deal comes as the U.S. military accelerates diversifying its AI suppliers following a bitter legal battle with Anthropic over safety guardrails.
Researchers just released Talkie, a 13B 'vintage' language model trained exclusively on 260B tokens of pre-1931 text. Because its worldview predates the web and digital computers, it sidesteps benchmark contamination entirely—revealing whether an AI can truly generalize and solve tasks like writing basic Python code with zero modern exposure.
Over 600 Google employees, including DeepMind researchers and more than 20 executives, just signed a petition demanding CEO Sundar Pichai ban the Pentagon from using Gemini in classified projects. Workers warned that secret military deployments leave them powerless to prevent potential harms behind closed doors. Why it matters: Google is currently discussing a classified cloud deal with the military, and internal resistance is once again threatening Pichai's defense ambitions.
A federal jury just rejected Elon Musk’s high-stakes lawsuit against Sam Altman and OpenAI after only two hours of deliberation, tossing the claims because the statute of limitations had run out. Musk cofounded OpenAI and sued in 2024, claiming the startup abandoned its founding nonprofit mission in favor of profits. Why it matters: After weeks of chaotic Silicon Valley drama on the stand, the trial ended in a total loss for Musk, leaving Altman’s grip on OpenAI fully intact.
Chinese regulators just ordered Meta to walk away from its planned $2B acquisition of AI agent startup Manus. Authorities prohibited foreign investment in the deal and demanded both sides withdraw. Manus builds agents that can plan and complete tasks independently. Why it matters: Meta had already begun deeply integrating the team, making any forced unwinding an operational and legal nightmare.
OpenAI just published five core principles to guide its pursuit of AGI, pledging to keep superintelligence decentralized rather than locked in the hands of a few labs. The framework covers democratization, empowerment, universal prosperity, resilience, and adaptability. To make AI universally accessible, OpenAI plans to aggressively expand global infrastructure while collaborating with governments on pathogen and cybersecurity risks.
Anthropic just detailed Claude's upgraded safeguards ahead of the 2026 US midterms and global elections. Models Opus 4.7 and Sonnet 4.6 scored over 95% on political bias benchmarks, and safety filters blocked nearly all attempts to run autonomous influence campaigns. Claude.ai will also serve banners routing voting queries to TurboVote.
OpenAI just released Privacy Filter, an open-weight model designed to detect and redact personally identifiable information in text. Built with 1.5B total and 50M active parameters, it supports up to 128,000 tokens of context and runs locally so sensitive data never leaves your device. It achieves a 97.43% F1 score across eight privacy categories.
Meta just started installing tracking software on U.S. employees' computers to record mouse movements, keystrokes, and occasional screenshots to train autonomous work agents. The company claims the data will not be used for performance reviews, but legal experts warn the practice brings gig-worker-level surveillance to white-collar desks.
YouTube just expanded its likeness detection tool to talent agencies and celebrities, helping them track down and remove AI face-swaps. The system works similarly to Content ID, and the kicker is that entertainers can now use it even if they don't have a YouTube channel of their own.
Anthropic just saw its high-stakes cybersecurity model, Mythos, accessed by unauthorized users through a third-party vendor environment. A group hunting unreleased models reportedly leveraged contractor access and guessed endpoints on the day of its announcement. Anthropic says it is investigating, with no direct breach found in its own systems.
OpenAI just expanded its Trusted Access for Cyber program, pledging $10 million in API credits to help defenders and open-source maintainers secure critical software. Industry giants like CrowdStrike, Cisco, and JPMorgan Chase are backing the push, while GPT-5.4-Cyber has been shared with US and UK safety institutes for evaluations. Why it matters: OpenAI wants to make sure defenders get their hands on frontier models before attackers do.
Anthropic just released Claude Opus 4.7 across its web apps and API, keeping pricing steady at $5 per million input tokens and $25 per million output tokens. The model brings notable improvements to complex software engineering, higher-resolution vision, and long-running autonomous workflows. It also debuts new safeguards to detect and block prohibited cybersecurity risks.
OpenAI just purchased access to Epoch AI’s FrontierMath verifiers to automatically check whether solutions generated by its models actually hold up. FrontierMath tests AI against extremely difficult mathematics problems, giving OpenAI automated, rigorous ground truth to evaluate advanced reasoning capabilities.
OpenAI just expanded its Trusted Access for Cyber program and introduced GPT-5.4-Cyber, a fine-tuned model built specifically for defensive cybersecurity work. The permissive variant lowers refusal boundaries on legitimate security tasks and adds binary reverse engineering to analyze compiled binaries for vulnerabilities. It allows verified defenders to leverage frontier capabilities without running into standard safety guardrails.
Anthropic just tested nine copies of Claude Opus 4.6 as automated researchers to tackle the challenge of weak-to-strong supervision. While human researchers achieved a PGR of 0.23 over seven days, the automated Claudes hit a PGR of 0.97 in five days for around $18,000. Why it matters: AI might soon accelerate its own safety research, though the method didn't yield significant gains when tested on Claude Sonnet 4 at production scale.
