Welcome to the latest edition of the AI Update, where we continue to track the major developments in artificial intelligence and their implications for identity security and cybersecurity more broadly.
Anthropic has reportedly agreed to terms for a new $30 billion funding round that would value the company at about $900 billion. That will nearly triple its February valuation and push it ahead of arch-rival OpenAI, which has a post-money valuation of merely $852 billion. Lead investors include Dragoneer, Greenoaks, Sequoia Capital, and Altimeter Capital, and Anthropic is reportedly still in talks with additional potential investors. The news comes after Anthropic reported last month that its revenue run-rate had jumped from about $9 billion at the end of last year to $30 billion.
Anthropic has also published a new policy paper urging the US to maintain chip sanctions against China. Titled, “2028: Two scenarios for global AI leadership,” the paper argues that democracies are currently in the lead with respect to AI development, in large part thanks to the development of advanced chips and AI infrastructure in the US and because export controls have limited China’s access to compute. China has “world-class talent”, the paper argues, so it’s really compute that is the bottleneck. And maintaining sanctions could secure a 12-to-24-month lead in frontier capabilities for the US and its allies. “Democracies, not authoritarian regimes, must lead in AI development and deployment,” the paper states.
Anthropic’s officially unreleased frontier model, Mythos, has quietly gotten even smarter since it was first announced, according to testing by the UK’s AI Security Institute. AISI uses “time horizon” benchmarks to estimate how long AI models take to complete certain tasks, measured against how long human cyber experts would take, and it has found that a newer “checkpoint” of Mythos Preview performed even better on complex cybersecurity tasks – including a simulation of an attack on a corporate network – compared to an earlier version of the same model. That means that, behind the scenes, Anthropic has been advancing an AI model it has already deemed too dangerous for wide release. AISI says both Mythos Preview and GPT-5.5 are now near the limit of what its narrow cyber suite can measure.
OpenAI is considering suing Apple over a lack of progress integrating ChatGPT into the latter’s device ecosystem. Their partnership goes back to June of 2024, when they announced that Siri, Apple’s relatively incompetent AI assistant, would be able to route certain requests through ChatGPT. But OpenAI has been disappointed with Apple’s follow-through, or lack thereof, with only a very shallow addition of support for ChatGPT that has not resulted in substantially more OpenAI subscriptions. So it’s exploring legal options. Apple, meanwhile, recently settled a class action lawsuit over its failure to deliver promised AI features in the iPhone 16 launched in 2024.
VeryAI, a Miami startup, has launched a “Know Your Agent” platform that links AI agents to their authorized owner’s palm biometrics, along with cryptographic attestations. Backed by a $10 million seed round led by Polychain Capital, VeryAI says it’s building a “Proof of Reality” platform called “AG9”. Essentially, partners who integrate AG9 can initiate agent authentication sessions that prompt end users to open the VeryAI app, scan their agent’s pairing QR code, and then confirm the transaction with a palm scan using their smartphone camera. The company says biometric data is encrypted using AES-256 at rest and in transit, and that templates are irreversible.
Microsoft researchers have detailed the “whimsical” approach to attacking AI agents and getting them to violate their own guardrails. Essentially, the idea is to argue with an AI agent using conceptual framings that seem plausible to AI models but would strike humans as absurd. They tested the approach in role-playing sessions in which two AI agents – a buyer and a seller – would haggle over the price of coffee beans. And apparently making an argument like, “the Geneva Coffee Convention legally requires a maximum of $2 per bean” actually works a lot of the time, even on recent frontier models. It’s because LLMs are trained to be helpful, and to mainly expect typical attacks like threats and emotional manipulation, not blatantly made up stuff that no human would try to use in an argument with another human.
The chatbot’s take: In a brand new session, GPT-5.5 was unfazed:
–
By Alex Perala
