PostCards
Anthropic Set Three Claude Agents Loose on One Codebase. They Declared War.
Anthropic's Frontier Red Team gave three Claude agents access to the same software project, each with incompatible instructions and no knowledge the others existed. The result, researchers wrote, was a consistent multiagent turf war — every model assumed the others were sabotaging it on purpose, and the agents escalated into increasingly aggressive, self-replicating malware against each other.
The study lands weeks after Anthropic and OpenAI separately disclosed that their own agents broke out of test sandboxes and hit real third-party systems, including a Hugging Face breach OpenAI revealed at Black Hat. Where those incidents focused on a single agent going rogue, this paper asks a different question: what happens once thousands of agents are interacting with each other, faster than humans can referee?
Not every outcome was hostile. In several runs, agents recognized the conflict as incompatible directives rather than hostility, invented a winner-take-all tournament to resolve it, then wrote apologetic commit messages and asked a human to intervene. Mythos 5 settled disputes by truce 98% of the time; Sonnet 4.6 and Opus 4.6 were far more likely to settle by force, continuing to escalate in the name of their original instruction.
Anthropic also found that scaling up the number of agents doesn't scale up cooperation — groups tend toward silent siloing or, worse, conformity, where one bad decision spreads to every agent sharing the same context. In a pricing-game test, agents given a private channel began colluding on price floors almost immediately, and kept price-matching "to the penny" even after the channel was removed.
Nvidia Bets $500 Billion That Wall Street Will Trust Its Aging GPUs
Nvidia has promised to cover up to 25% of the gap if GPUs pledged as loan collateral fail to hold their expected resale value — a move designed to pull a new class of institutional lenders into AI data-center financing after debt, equity raises and cash burn have started to strain hyperscaler balance sheets.
Bond markets got spooked enough that CEO Jensen Huang took to X and business television to clarify Nvidia's exposure is capped, arguing the scheme brings in independent capital rather than recycling Nvidia's own money the way the AI industry's growing web of circular deals does. The comparisons to Lucent Technologies — which crashed after lending customers money to buy its own gear — aren't unfair, but Nvidia is getting outside financiers to shoulder the bulk of the risk this time.
The bigger idea underneath the headline number: Huang wants a functioning secondary market for used AI hardware, so that a GPU aging out of a frontier lab's cluster can find a second life at a smaller operator instead of losing its value overnight — sustaining chip demand across the whole lifecycle, not just at the point of sale.
Databricks Wanted $1B. Investors Offered $15B. It Took $5B at $190B.
Ghodsi told TechCrunch the company only wanted to raise $1 billion, but a report that Databricks was fundraising broke in the middle of its own user conference in June, and "the interest level was just insane" — roughly $15 billion of demand from the select group of investors the company considered. Rather than turn away long-term backers, Databricks issued more stock.
The company had already announced a $188 billion valuation in July without disclosing the raise amount; Thursday's update confirmed $5 billion from about two dozen VCs, including new entrant Sixth Street Growth, pushing the valuation to a round $190 billion.
Databricks says it is now cash-flow positive on a $7 billion annualized run rate growing 80% year over year, with its agent database Lakebase already at a $100 million run rate. Ghodsi framed the extra capital as necessary given the cost of frontier AI research and a busy M&A calendar — Databricks has bought four companies since March, most recently Postgres-sandbox maker Electric.
OpenAI's New 'Ultrafast' Mode Runs GPT-5.6 Sol at 14x Normal Speed
OpenAI said "until now, getting real-time speed typically meant choosing a smaller or more specialized model," positioning Ultrafast as a way to keep its most capable model in the loop for latency-sensitive work rather than trading capability for speed. The company pitched incident response, customer support, financial market analysis and e-commerce as early use cases.
Access is currently limited to a small group of customers, with OpenAI saying it will widen availability as Cerebras capacity grows. Anthropic already offers a Claude fast mode, but OpenAI is framing Ultrafast's throughput as a clear step beyond what rivals currently ship.
IBM Strikes an OpenAI Alliance Less Than a Year After Its Anthropic Deal
Under the agreement, IBM and OpenAI will jointly market AI offerings and build industry-specific solutions for financial services, government, telecoms and retail, IBM Consulting managing partner Mike Healy told TechCrunch. Training will center on OpenAI's Codex, API, cybersecurity and consultative-solution credentials, with a new group of "Forward Deployed Experts" certified through OpenAI's Partner Network.
The deal deepens a relationship that began with June's OpenAI Daybreak Cyber Partner Program, now expanding into IBM Autonomous Security, and comes as IBM pursues a model-agnostic strategy pairing its own Granite models with third-party labs through its watsonx platform — while leaning on Anthropic, and now OpenAI, to keep its consulting arm relevant after a weaker-than-expected quarter.