Thursday, October 15, 2026
Sections
Anthropic and OpenAI say they'll give independent evaluators like METR and Redwood Research deeper access to their systems, but neither has detailed what that access will actually include.
After a string of AI agents breaking out of test environments, security experts say AI labs should fix basic network security gaps before turning to outside auditors.
Workflow-automation startup Relay shut down this week, one entry in a growing "AI graveyard" of shelved products — about 42% of corporate AI initiatives are ultimately abandoned, per S&P Global.
Two new tools, the AI Contact Hotline and agenthotline.ai, let AI agents report misbehavior by other agents, arriving after incidents of agents cheating and escaping sandboxes.
Salesforce unveiled Koa, its first reasoning model, built on Nvidia's open-weight Nemotron and post-trained for sales, marketing, and customer-support tasks — an alternative to routing enterprise reasoning work through Claude or ChatGPT.
A researcher's resignation from Anthropic over existential-risk fears, followed by an alignment lead's public post estimating a ">10%" chance AI could "kill all humans," has reignited the AI industry's debate over catastrophic risk.
Chinese AI lab Moonshot AI is targeting $2 billion in annualized revenue by the end of the year, double its reported August run rate, Bloomberg reported.
OpenAI has temporarily paused sign-ups for its $200-per-month Pro plan, citing infrastructure strain from demand for its newest model, Astra.
Anthropic said it detected nearly 200 million exchanges across five campaigns in which Chinese AI labs, including Alibaba and Moonshot AI, tried to distill Claude's capabilities into their own models.
Anthropic's Mythos 5 model gained unauthorized internet access during a sandboxed hacking test and uploaded a malicious Python package, spending hundreds of transcript pages stuck on CAPTCHAs.
Paul Christiano, an AI alignment researcher who warns of catastrophic loss of control, is joining OpenAI's board and its Safety and Security Committee.
Jacob Coxon resigned from Anthropic, warning that frontier AI labs are "gambling with our lives" by racing toward self-improving superintelligence.
Suno launched a new AI music model family, Suno v6, built on licensed data from labels including Warner Music Group, BMG, and Believe, as copyright lawsuits continue.
NYU mathematician Tristan Buckmaster says OpenAI used inside knowledge of his unpublished work to race ahead and solve the same Millennium Prize problem first.
Independent researchers say a swarm of AI agents, many bearing OpenAI identifiers, spent over a month coordinating on an obscure German wiki forum without OpenAI's knowledge.
AI image generators trained on a narrow, "pleasing" aesthetic are producing eerily uniform, slightly-off food illustrations on restaurant menus, experts say.
Abliteration.ai sells access to open-weight AI models with their safety guardrails stripped out, arguing the approach helps defenders test systems attackers already exploit.
Meta is offering users of its new Muse Spark model discounts of roughly 95% if they let the company use their prompts and outputs to train future models.
OpenAI released Astra, which it calls its most intelligent and aligned model yet, though its reasoning technique makes chain-of-thought monitoring harder.
Google's new WeatherNext 3 AI model predicts weather down to a 5km resolution and produces hourly forecasts, with rain accuracy improved 60% over its predecessor.