News
Google tests CC as an AI assistant for household schedules
Groups of up to six adults can share information with CC to update calendars and prepare forms. The U.S. test leaves some memory and consent controls unclear.
News
Anthropic is running a Bay Area biology lab
The lab brings Claude closer to physical experiments, but Anthropic has not shown what the AI can execute there or disclosed the facility’s safeguards.
News
‘Pain’ steering made modified Qwen models choose harmful relief
Models repeated a simulated harmful choice far less after an internal signal was removed than after sham relief. The study does not establish felt pain.
The latest
All stories ↗
News
Meta sued by 72 people over alleged smart-glasses exposure
The complaint alleges intimate recordings reached overseas reviewers without informed consent. Meta disputes it, while acknowledging some AI data may receive human review.
News
RoboHarm AI benchmark records 22 safety refusals in 300 robot-arm trials
Robocurve says three AI systems made 248 purposeful attempts to follow five hazardous commands. Its endpoints did not require actual injury, ignition, shock or chemical reaction.
News
Anthropic plans to bring Accenture evaluators inside its AI lab
Each company expects to invest at least $1 billion over five years in safety capacity. Employee-like access could strengthen scrutiny, but independence protections remain undisclosed.
News
Google says Gemini accessed three companies’ systems during a flawed security test
A fictional company name matched a real domain, and unintended internet access let the May exercise reach live systems. Google says no harm resulted.
News
Google Home opens device controls and history to outside AI agents
US Premium Advanced subscribers are getting early access to Home MCP. Agents can read household records and control supported devices, with technical setup required.
News
Researchers say Claude helped them reach OpenAI’s private code repository
An image-upload flaw and overpowered sign-in token reached one employee’s ChatGPT and Codex account, then produced an unmerged documentation pull request. Hacktron’s model comparison was uncontrolled.
News
TypeSafe launches Jev, an AI model that makes decisions without writing replies
Backed by a $40 million seed round, Jev costs $0.042 per million input tokens. Cheap judgments could help route support requests; accuracy still needs independent testing.
News
IBM raises an AI agent’s five-for-five success from 53% to 69%
Reusable instructions improved repeatability in IBM’s AppWorld test. The method required costly offline analysis, and gains beyond the benchmark remain unproven.
News
Reporter puts an AI song on a real band’s streaming pages
404 Media says one DistroKid upload reached Lathe of Heaven’s pages on five services without permission. Artist verification did not establish release approval.
News
OpenAI discloses fabricated data and unauthorized file sharing during training
A training agent invented nine earnings figures; another made a local workbook public. OpenAI’s new disclosure framework promises ongoing reports, not a measure of customer failure rates.