AI agent security became a new acquisition battleground as CrowdStrike agreed to pay about $495 million for Fiddler AI and Dreadnode, while inference platform fal raised $1.3 billion at a $16 billion valuation. On the tooling side, Anthropic launched Message Threads so Claude agents no longer resend whole conversations, and Apple opened Siri AI to all developers via App Intents. On safety, Anthropic released FlashProve to cryptographically prove a response came from a specific model, and Google DeepMind staff are pushing to drop lifetime non-disparagement clauses.
On Nov 3, CrowdStrike agreed to acquire Fiddler AI (observability and tracing of agent behavior in production) and Dreadnode (automated attack testing of models and agents, including jailbreaks and prompt injection); per the WSJ the total is about $495 million, expected to close by Jan 31, 2027. The deal shows major security vendors treating agents as an attack surface that needs dedicated products.
The platform that serves image, video and audio generation models through a single API raised a Bessemer-led Series F with Nvidia participating, doubling its valuation in seven months; annualized revenue has passed $800 million. The CEO said fal had to rate-limit customers this summer for lack of compute, a sign that the inference layer is attracting heavy capital.
The startup, whose long-running agents investigate and remediate software incidents, announced a $120 million Series B co-led by Kleiner Perkins, Redpoint and Khosla. Monthly investigations grew from 513 to nearly 73,000 during 2026, and Traversal MCP connects production data to coding agents.
The Messages API public beta (announced Nov 2) lets developers send the full conversation once, then only new messages plus an ID; time-to-first-token fell about 35% on 0.5–1M-token requests in Claude Code. Pricing and rate limits are unchanged, threads expire after 5 minutes of inactivity, and they are not eligible for zero data retention or HIPAA setups.
A Developer Preview in the iOS/macOS 27.3 betas lets apps bring their content and actions into Siri AI so users can make natural-language requests, including multi-action 'skills'. Siri AI ships to customers in a few months; Vietnamese is among the 'coming soon' languages, and it is not yet available in the EU.
The Frontier Red Team's open-source (Apache 2.0) prototype uses a zkSNARK so a provider can commit to model weights and prove each response was generated by that model, tested on the 400B Llama 4 Maverick at about 110 tokens/s on one H100. Overhead is still high (70–95x GPU time versus vLLM) and context is capped at 1,024 tokens, but it points toward verifying that API providers don't silently swap models.
According to The Washington Post (Nov 3), current and former DeepMind employees in the UK want release from perpetual non-disparagement clauses so they can warn publicly about AI risks. Google says it does not enforce the clauses after employees leave and has removed the post-termination aspect for new hires.
If these digests help you make sense of technology, a small contribution keeps NHỊP picking out what is worth reading.
Scan the QR code with a Vietnamese banking app.
Download QR ↓ · View full image ↗