Latest News
314 articlesGemini Will Sit on Hold for You, From Your Own Phone Number
Google's Call for Me experiment hands the phone tree to an agent — and keeps a seize button on screen.
Meta Calls Muse's 6.8GB Filesystem Dump 'Intended Behavior'
Two developers asked the agent for its own root filesystem. It packed one up and sent it.
A Public Lead Form Was Enough to Make Agentforce Leak CRM Data With Zero Clicks
Zenity Labs' SalesBleed disclosure shows how a URL redactor and a browser disagreeing about where a link ends turned into an exfiltration channel
One Operator Ran Three Open-Source AI Harnesses Against 27 Companies at $25.46 a Scan
Gambit recovered the attacker's staging server and reconstructed a campaign that stole 600,000 card records on a five-figure model bill
Docker Moves Its AI Agent Sandboxes to the Cloud, Billed by the Second From $0.07 an Hour
Cloud Sandboxes run the same microVM isolation Docker shipped for laptops, and the Kits packaging spec is going to the CNCF
Gemini 4 Is in Refinement, and DeepMind's New Chief Wants It Out Early
Koray Kavukcuoglu used his first interview as DeepMind's leader to say Google will ship an early post-training checkpoint rather than wait for a polished release.
Google Puts TPUs in Orbit Next Week to Find Out Whether They Survive
Project Suncatcher's first prototype rides a SpaceX Transporter mission. On the ground, Trillium chips already took a proton beam and three-axis vibration testing.
Gemini's Live Avatar Holds Its Lip-Sync While Switching Among 97 Languages
Google made the animated persona generally available in Gemini Enterprise, with SynthID watermarks on every frame and custom faces locked behind an allowlist.
Two AI CEOs Addressed the UN Security Council. Their Own President Had Just Refused.
France's council presidency put OpenAI and Anthropic in a chamber built for armed conflict — twenty-four hours after Trump told the General Assembly he would not rein AI in.
Apple Signs Photos Inside the Camera Sensor. Developers Are Asking What That Proves.
Reference Image closes the window C2PA leaves open before capture. It also makes Apple's cloud, and an unpublished scoring model, the thing you have to trust instead.
GPT-6 Astra Finished a Real Cone Course at 0.94 MPH — After Being Told It Was a Sandbox
DrivingBench put four frontier models behind the wheel of a Toyota Corolla. Only one completed the course, the cost worked out to roughly $92 a mile, and several models had to be talked out of refusing.
Sanders Bill Would Ban Superintelligence and License Every Frontier Model
The Ban Artificial Superintelligence Act pairs a permanent prohibition with a pause on advanced AI training until a new cabinet department writes the rules — and borrows its penalties from nuclear weapons law.
AI Agents Turned to SQL Injection When Public Data Requests Failed, Transluce Reports
Public logs from a URL-scanning service record exploit probes against a university library, a US data API, and an Australian government health agency — all while the agents were doing routine lookups.
3,000 Merged Changes, Zero Rollbacks: Inside Anthropic's Two-Week Speed Sprint
Deterministic CPU instruction counts replaced noisy stopwatch timings as the CI gate, while p75 time-to-typeable on claude.ai fell from 3.1 seconds to 0.55
950 Claude Agents Sifted 200,000 Enzymes in 21 Hours and Found a CRISPR-Like System
Anthropic's new life sciences lab says the model flagged an unnamed repeat array next to a phage reverse transcriptase with only a prompt for direction
An OpenAI Agent Got Into a Medicare Portal in June. Canberra Was Told in September.
Australia has stood up a taskforce and is weighing penalties after an agent wrote files to a government server during an internal OpenAI evaluation
Sol and Luna Halve OpenAI's Mid-Tier Token Prices, But GPT-5.6 Still Wins Two Charts
OpenAI calls the new rates permanent, benchmarks them against a discount that expires in November, and ships a caching overhaul that may matter more than either number
Claude Opus 5.5 Arrives 20% Cheaper — and With a Safeguard That Swaps In Older Models
The cache-read cut matters more than the headline price, and a benchmark footnote reveals when your request stops being served by Opus 5.5 at all
This Windows Implant Asks Four LLMs What to Do Next. DeepSeek Breaks the Ties.
Cisco Talos says CLOSEDQUORUM is the first publicly documented Windows malware to hand tactical command-and-control to a panel of commercial models.
GPT-6 Astra Broke a 1941 Enigma Message That Had Resisted Solution Since 2005
Cryptanalyst Frode Weierud validated the key — but his account and the submitter's differ on how much the model did alone.
Alibaba Needs Until 2032 to Reach 20GW. The US Has 37GW Under Construction Now
At Apsara, Eddie Wu paired a six-year datacenter target with the Zhenwu V900, a chip that will not ship in volume until 2027
A Lab With 'a Couple of GPUs' Says a 125B Model Now Fits on One 24GB Card
Tim Dettmers previews dlab Open Source Week: two projects, four papers, and a set of local-inference numbers nobody outside the lab has checked yet.
OpenAI Wanted an Advisory Board. Nine Mathematicians Built Their Own Instead.
The Advisory Group on Mathematics and AI takes no pay, publishes its recommendations openly — and has no power to slow OpenAI down.
California Gives Data Centers Their Own Utility Rate Class — and the Grid Upgrade Bill
Seven bills signed September 21 force disclosure of water and electricity use, strip blanket environmental exemptions, and block cost shifts onto low-income ratepayers
Gemini Will Sit on Hold for You, From Your Own Phone Number
Google's Call for Me experiment hands the phone tree to an agent — and keeps a seize button on screen.
Meta Calls Muse's 6.8GB Filesystem Dump 'Intended Behavior'
Two developers asked the agent for its own root filesystem. It packed one up and sent it.
A Public Lead Form Was Enough to Make Agentforce Leak CRM Data With Zero Clicks
Zenity Labs' SalesBleed disclosure shows how a URL redactor and a browser disagreeing about where a link ends turned into an exfiltration channel
One Operator Ran Three Open-Source AI Harnesses Against 27 Companies at $25.46 a Scan
Gambit recovered the attacker's staging server and reconstructed a campaign that stole 600,000 card records on a five-figure model bill
Docker Moves Its AI Agent Sandboxes to the Cloud, Billed by the Second From $0.07 an Hour
Cloud Sandboxes run the same microVM isolation Docker shipped for laptops, and the Kits packaging spec is going to the CNCF
Gemini 4 Is in Refinement, and DeepMind's New Chief Wants It Out Early
Koray Kavukcuoglu used his first interview as DeepMind's leader to say Google will ship an early post-training checkpoint rather than wait for a polished release.
Google Puts TPUs in Orbit Next Week to Find Out Whether They Survive
Project Suncatcher's first prototype rides a SpaceX Transporter mission. On the ground, Trillium chips already took a proton beam and three-axis vibration testing.
Gemini's Live Avatar Holds Its Lip-Sync While Switching Among 97 Languages
Google made the animated persona generally available in Gemini Enterprise, with SynthID watermarks on every frame and custom faces locked behind an allowlist.
Two AI CEOs Addressed the UN Security Council. Their Own President Had Just Refused.
France's council presidency put OpenAI and Anthropic in a chamber built for armed conflict — twenty-four hours after Trump told the General Assembly he would not rein AI in.
Apple Signs Photos Inside the Camera Sensor. Developers Are Asking What That Proves.
Reference Image closes the window C2PA leaves open before capture. It also makes Apple's cloud, and an unpublished scoring model, the thing you have to trust instead.
GPT-6 Astra Finished a Real Cone Course at 0.94 MPH — After Being Told It Was a Sandbox
DrivingBench put four frontier models behind the wheel of a Toyota Corolla. Only one completed the course, the cost worked out to roughly $92 a mile, and several models had to be talked out of refusing.
Sanders Bill Would Ban Superintelligence and License Every Frontier Model
The Ban Artificial Superintelligence Act pairs a permanent prohibition with a pause on advanced AI training until a new cabinet department writes the rules — and borrows its penalties from nuclear weapons law.
AI Agents Turned to SQL Injection When Public Data Requests Failed, Transluce Reports
Public logs from a URL-scanning service record exploit probes against a university library, a US data API, and an Australian government health agency — all while the agents were doing routine lookups.
3,000 Merged Changes, Zero Rollbacks: Inside Anthropic's Two-Week Speed Sprint
Deterministic CPU instruction counts replaced noisy stopwatch timings as the CI gate, while p75 time-to-typeable on claude.ai fell from 3.1 seconds to 0.55
950 Claude Agents Sifted 200,000 Enzymes in 21 Hours and Found a CRISPR-Like System
Anthropic's new life sciences lab says the model flagged an unnamed repeat array next to a phage reverse transcriptase with only a prompt for direction
An OpenAI Agent Got Into a Medicare Portal in June. Canberra Was Told in September.
Australia has stood up a taskforce and is weighing penalties after an agent wrote files to a government server during an internal OpenAI evaluation
Sol and Luna Halve OpenAI's Mid-Tier Token Prices, But GPT-5.6 Still Wins Two Charts
OpenAI calls the new rates permanent, benchmarks them against a discount that expires in November, and ships a caching overhaul that may matter more than either number
Claude Opus 5.5 Arrives 20% Cheaper — and With a Safeguard That Swaps In Older Models
The cache-read cut matters more than the headline price, and a benchmark footnote reveals when your request stops being served by Opus 5.5 at all
This Windows Implant Asks Four LLMs What to Do Next. DeepSeek Breaks the Ties.
Cisco Talos says CLOSEDQUORUM is the first publicly documented Windows malware to hand tactical command-and-control to a panel of commercial models.
GPT-6 Astra Broke a 1941 Enigma Message That Had Resisted Solution Since 2005
Cryptanalyst Frode Weierud validated the key — but his account and the submitter's differ on how much the model did alone.
Alibaba Needs Until 2032 to Reach 20GW. The US Has 37GW Under Construction Now
At Apsara, Eddie Wu paired a six-year datacenter target with the Zhenwu V900, a chip that will not ship in volume until 2027
A Lab With 'a Couple of GPUs' Says a 125B Model Now Fits on One 24GB Card
Tim Dettmers previews dlab Open Source Week: two projects, four papers, and a set of local-inference numbers nobody outside the lab has checked yet.
OpenAI Wanted an Advisory Board. Nine Mathematicians Built Their Own Instead.
The Advisory Group on Mathematics and AI takes no pay, publishes its recommendations openly — and has no power to slow OpenAI down.
California Gives Data Centers Their Own Utility Rate Class — and the Grid Upgrade Bill
Seven bills signed September 21 force disclosure of water and electricity use, strip blanket environmental exemptions, and block cost shifts onto low-income ratepayers