In a matter of weeks, Claude generated roughly 180 million lines of machine-checked proofs for 123 open-source libraries and fixed hundreds of vulnerabilities. An audit of OpenAI’s internal agents found that context compaction left about 10% of eligible internal Codex traffic unmonitored. Microsoft’s MAI-DxO achieved 83% diagnostic accuracy across 304 complex cases; its next test will assess clinical performance through direct patient interaction. Modal raised $750 million at a $15.75 billion valuation, with its annualized revenue run rate surpassing $1 billion, including $300 million from Sandboxes.
Proving that specific flaws are absent does not mean misaligned agent behavior can be caught during execution.
Anthropic’s Frontier Red Team partnered with Theorem Labs to use Claude to verify security properties across 239 entry points in 123 open-source libraries in a matter of weeks. The effort generated roughly 180 million lines of machine-checked proofs and fixed hundreds of vulnerabilities. Formal verification can prove that code is free of specific types of flaws, but it has historically been limited to small codebases or required substantial expert time.
Read original →Two auditors from METR and Redwood Research spent three weeks on-site at OpenAI assessing its system for monitoring misaligned behavior by internal agents. They found that agents could delay monitoring indefinitely by not handing back control, and that one vulnerability could let hundreds of recent actions escape monitoring. Context compaction also left about 10% of eligible internal Codex traffic unmonitored.
Read original →🔍 Analysis: Even if Claude can provide proofs of specific security properties for 123 libraries, that does not eliminate the monitoring blind spots involving control handoffs and context compaction identified in OpenAI’s audit. Code properties and runtime monitoring coverage are two separate things to validate. For companies deploying agents, security validation cannot stop at code-checking results. It also needs to establish whether monitoring depends on agents voluntarily handing back control, and whether compacted activity records are available for auditing.
MAI-DxO’s next test moves from interactive case evaluations to interactions with real patients.
Microsoft’s diagnostic system, MAI-DxO, achieved 83% accuracy in an evaluation involving 304 complex NEJM cases, compared with 36% for doctors using reference materials and 22% without them. Rather than delivering an answer in one shot, it has multiple virtual doctors discuss the case over several rounds, progressively selecting questions and tests while weighing costs. The evaluation used interactive cases. Microsoft will conduct a prospective study with Beth Israel Deaconess Medical Center in Boston, allowing patients to interact directly with the system to assess its real-world clinical performance.
Read original →Modal demonstrates paid demand at the execution layer, while Amazon continues investing to address supply shortfalls.
Modal completed a $750 million Series D led by Accel at a post-money valuation of $15.75 billion, with its annualized revenue run rate surpassing $1 billion. Just five months after its Series C, usage has tripled and grown more than 15-fold year over year. Sandboxes, launched in 2023, has reached an annualized revenue run rate of $300 million on its own. Nearly 3 billion sandboxes have been launched to date, and the product can scale to a million concurrent instances for uses including background agents and reinforcement learning.
Read original →Amazon expects capital expenditure of around $220 billion in 2026. CEO Andy Jassy said that even with investment at this scale, the company still cannot meet a substantial share of AI demand. He believes this infrastructure expansion could eventually help AWS reach $1 trillion in annualized revenue, while lower inference costs could also encourage broader enterprise adoption of AI.
Read original →🔍 Analysis: Modal’s Sandboxes now accounts for roughly 30% of its annualized revenue run rate, showing that agent infrastructure revenue comes not only from underlying compute but also from isolated execution and concurrency scheduling. Amazon’s projected $220 billion in capital expenditure, meanwhile, reflects expansion of the underlying supply. These are not directly comparable growth metrics: the former is already backed by product revenue, while returns on the latter still depend on how quickly demand absorbs the added capacity.
Governance of workflow-knowledge standards and procurement channels for model services are becoming separate matters.
Anthropic has donated Agent Skills to AAIF under the Linux Foundation, with more than 70 agent platforms already integrating the standard. Originating in Claude Code, it uses skill folders to supply business-process knowledge. The transfer includes the name, logo, GitHub organization, and domain to preserve vendor neutrality. AAIF membership is not required to use or contribute to the standard.
Read original →Anthropic’s hosted Claude service on Google Cloud is now generally available, supporting existing Google Cloud IAM credentials, consolidated billing, and committed-spend credits. The new offering receives new features and beta features on the same day as the native Claude platform, and already supports code execution, the Skills API, and Opus 5.5 fast mode. It currently offers US and global endpoints, with Anthropic serving as the data processor. Customers who need Google to act as the data processor or require strict regional data residency can still choose the Google-hosted version.
Read original →🔍 Analysis: The transfer of Agent Skills to AAIF, with more than 70 platforms already integrating it, reduces the risk of tying business-process knowledge to a single vendor. Claude’s Google Cloud service integration with IAM, billing, and committed spend addresses a different issue: friction in enterprise procurement and integration. For enterprises, skill formats can be managed under a neutral standard, but model services still require a separate choice. Same-day access to new features and strict regional data residency are not both available through the same hosting arrangement.
MiniMax reduces the deployment burden of a general-purpose model, while Carbon-X redesigns representations for DNA tasks.
MiniMax has released the weights for M3.1 Flash, which has 428 billion total parameters and activates 23 billion per token. The model supports a 1-million-token context window and visual input. vLLM provided deployment support on day one, using native NVFP4 weights and KV cache to reduce the checkpoint size by 44.6% to 250.1GB.
Read original →Hugging Face has released Carbon-X, which combines DNA and text and was mid-trained from Qwen3-14B on 300 billion mixed tokens. The team says its DNA benchmark score is less than one point behind GPT-6.1 Sol, with inference costs more than 1,000 times lower. The model uses 6-mer DNA tokenization, producing roughly one-third as many tokens as BPE. It is released under the Apache 2.0 license, can run on a laptop, and excluded human and animal viruses from training.
Read original →🔍 Analysis: M3.1 Flash uses native NVFP4 to compress weights and the KV cache, while Carbon-X uses 6-mer tokenization to shorten DNA representations. One reduces deployment overhead; the other changes the input representation for a specific task, so their cost figures are not directly comparable. For developers, open weights do not imply the same hardware requirements: a 250.1GB checkpoint and a model that can run on a laptop represent different deployment conditions. Model selection should first distinguish general-purpose multimodal needs from DNA-specific tasks, rather than simply comparing parameter counts or benchmark scores.