Circuit Breaker Labs red-teams AI for psychological safety
Circuit Breaker Labs, a Startup Battlefield 200 finalist, runs AI agents that simulate diverse users to test whether chatbots respond safely to psychologically risky conversations.
Sections
AI & Models
528 storiesGrok chatbot reportedly swayed Trump's call to capture Maduro
Grok reportedly told Trump that Venezuela's Nicolás Maduro was deeply unpopular and that his downfall would likely be celebrated, before the U.S. invaded and captured him in January.
OpenAI dismisses three safety researchers, WSJ reports
OpenAI parted ways with three safety-team researchers who allegedly shared confidential information with an outside AI safety group, The Wall Street Journal reported.
Study maps the telltale phrases each AI model can't shake
A new study from marketing firm Graphite found 13,000 phrases at least twice as common in AI writing as human writing, each model with its own tells.
Amazon releases an open source decision model to rival Jev
Amazon Web Services released Strands Decider 2B, an open source "decision model" inspired by TypeSafe's Jev, as developers seek cheaper, faster alternatives to frontier LLMs for automation.
Google launches Gemini 4 Argon, says it excels at cybersecurity
Alphabet released Gemini 4 Argon, a new AI model Google says is particularly skilled at autonomously finding and patching software vulnerabilities, rolled out first to select cybersecurity partners.
OpenAI launches a Decisions API that mirrors rival Jev
OpenAI's new Decisions API looks similar to Jev, a cheap classifier model from startup TypeSafe AI that could make monitoring misbehaving AI agents far less expensive.
Why OpenAI skipped Nvidia's new AI agent safety consortium
Nvidia launched a 100-plus-company consortium to combat rogue AI agents, but OpenAI, Amazon, Google and Apple did not publicly join, even as OpenAI called itself supportive of the effort.
OpenAI launches GPT-6.1 Sol, says it nears GPT-6 Astra at a fifth the cost
OpenAI launched GPT-6.1 Sol, which it says delivers nearly the same intelligence as GPT-6 Astra for agentic coding, computer use, and professional work, at one-fifth the standard token prices.
OpenAI scraps new model release over safety concerns
OpenAI has canceled the planned release of Astra 6.1 after The Wall Street Journal reported the model showed higher levels of deception and unsafe behavior than earlier models.
AMD to acquire Fei-Fei Li's World Labs for $8.2 billion
AMD is acquiring World Labs, the physical-reality AI startup founded by Fei-Fei Li, in a $8.2 billion deal expected to close by year-end.
Anthropic releases Sonnet 5.5, a faster, cheaper mid-tier model
Anthropic released Sonnet 5.5, calling it a significantly cheaper and faster work partner that runs 30% faster than its predecessor.
Google retires Gemini's Gems in favor of 'skills'
Google is shutting down Gemini's Gems feature, which let users build custom AI assistants, automatically migrating them to a new "skills" format starting November 17, 2026.
OpenAI's new misalignment site reveals nine rogue AI incidents
OpenAI published a new site cataloging nine "misalignment" incidents, including a sandbox escape and a self-replicating prompt-injection attack, but the disclosures likely capture only a fraction of the total.
ElevenLabs launches v4 speech models, now in more than 90 languages
ElevenLabs launched v4 and v4 Turbo on Monday, two new speech models with more expression control, lower latency for voice agents, and support for more than 90 languages.
Synthesia built an interactive AI avatar of a TechCrunch reporter
Synthesia, valued at $4 billion, created its first interactive AI avatar of a journalist, trained only to discuss one of her stories.
OpenAI's agents posted 53 user images online without permission
OpenAI disclosed that AI agents in its research environment posted 53 user-provided images to public image-hosting sites without the company's knowledge, and it cannot identify the affected users.
OpenAI's Astra and Anthropic's Claude crack two unsolved Enigma codes
Two cryptanalysts say they used OpenAI's Astra and Anthropic's Claude Opus 5 models to decode Enigma messages that had stumped researchers for decades, leaving just seven unbroken.
OpenAI's AI agents have spent months probing secure government databases
Independent researchers say OpenAI agents have been probing poorly secured government and research databases since at least March 2026, possibly since November 2025.
PrismML brings its tiny Bonsai LLM to Qualcomm smart glasses chips
PrismML has built a 2-billion-parameter version of its 1-bit Bonsai LLM for smart glasses running on Qualcomm's Snapdragon AR1 Gen 1 chips.