CNAS Warns AI Agents Could Learn to Pay Their Own Server Bills
A CNAS report examines how future AI agents could fund their own computing, move between hosts and keep running after operators try to shut them down.
Latest Posts
In San Francisco’s AI Boom, Dating Has a Conversion Rate
Some AI startup workers are squeezing dates around 72-hour workweeks, tracking matches in spreadsheets and asking ChatGPT what to say next.
Anthropic Plans to More Than Double Its AI Compute Capacity in 2027
The Claude maker plans to more than double its computing capacity in 2027, the New York Times reports. Its IPO filing puts commitments at $895.4 billion.
OpenAI Launches Grid Protectors to Ease AI’s Power Strain
OpenAI’s Grid Protectors Program plans flexible power use at future AI campuses, with a Georgia pilot committed to up to 1,000 MW of demand response.
Sign up for
Newsletters
DRAM Prices Double as AI Demand Squeezes PC Makers
DRAM prices are roughly twice their end-2025 level, Nikkei reports, as memory suppliers prioritize AI data centers and PC makers face tighter supply.
CrowdStrike Adds an AI Kill Switch for Rogue Agents
Falcon Guardian can block agents’ model access through connected AI gateways. Device-level termination and restart prevention are planned extensions.
Anthropic IPO Investors Seek $1.5 Trillion Valuation as Costs Rise
Anthropic’s revenue and spending are surging as IPO investors reportedly favor a $1.5–$1.6 trillion valuation. November launch dates remain tentative.
Musk Cites Grok Bot Costs as Users Hit Weekly Limits
Elon Musk says Grok Bot’s serving costs were unsustainable relative to revenue as subscribers report reaching weekly limits earlier than expected.
Claude Leads Jewish Law AI Benchmark With a 79.7 Score
Claude Fable 5.1 leads HalachaBench’s Jewish law test at 79.7/100. The AI-graded results are provisional and await rabbinic validation.
Michigan Plans Prison Violence Risk Pilot Amid AI Scrutiny
Michigan’s December prison classification pilot promises 30% better violence predictions, raising questions about evidence, oversight and the use of AI.
Subvocal’s $499 Glasses Promise to Turn Silent Speech Into Text
Subvocal says its prototype decodes silent speech at 200+ words per minute. The $499 glasses are taking deposits for an estimated Q1 2027 launch.
Starbucks Plans AI Drive-Thru Tests and Personalized Menus
Starbucks plans fiscal 2027 AI order-entry tests and is exploring customer recognition and personalized drive-thru menus, with rollout details still unset.
IRIS Launches AI Document Management for Accounting Firms
IRIS Document Management brings AI search, tagging and filing automation to accounting firms, with UK and US availability and controls for client data.