$ today_in_10_seconds
Agent identity stopped being a whitepaper topic this week. CrowdStrike shipped runtime discovery and kill-switches for agents on the endpoint, Visa, Mastercard and Ant agreed to recognize each other's agent credentials, and Intel put encryption for agent-to-agent traffic into silicon. All three are the same admission: nobody can currently tell you which agents are running, who deployed them, or what they touched.
$ lead_story
CrowdStrike Falcon Guardian secures autonomous AI agents
CrowdStrike used Fal.Con 2026 to launch Falcon Guardian, extending the Falcon platform to discover, monitor, control and detect autonomous agents across endpoints, cloud, SaaS and browsers. The mechanics matter more than the category name: the Falcon sensor inventories known and shadow agents on Windows and macOS along with who deployed them, runtime visibility builds a causal chain from prompt to identity to tool call to downstream system action, and access controls decide which agents are permitted to run at all. CPO AJ Shipley's framing is the honest version of why this exists — agents don't type, they act — and policy documents cannot intercept a process that is already executing. The AI Gateway piece, which would apply the same policy to every AI call including MCP connections, is still pre-beta with GA targeted next quarter.
// why_it_matters: Your agent's execution graph is now an audit artifact someone else sells — build it yourself or buy it, but stop pretending the logs you have today constitute one.
$ also_worth_attention
Intel secures AI agents at the edge for autonomous ride performance
Atsign and Intel say they've cleared the encryption bottleneck on autonomous edge agents, claiming roughly 88x higher throughput for hardware-accelerated, end-to-end encrypted Agent2Agent traffic, pitched at robotaxis and smart-transport infrastructure. Treat an 8700% number from a joint vendor brief as directional until someone reproduces it. The signal worth keeping is that A2A encryption overhead is now considered a silicon problem, which only happens when agents are genuinely talking to each other in production.
link →Google: attackers used AI agents to steal credentials in under six hours
GTIG's "From Prompting to Autonomy" report covers Q2 activity and its headline finding is how little human involvement is left — adversaries handing multistep decisions to models and shrinking the defender's reaction window. The supply-chain half is uglier: UNC6780 ran compromises across PyPI, npm and Docker Hub, publishing trojanized forks of MCP servers and poisoning repos that coding assistants clone, with a credential stealer that hides in .claude and .cursor directories where AI tooling treats it as normal project clutter. Your agent's dependency tree is now an attack surface with a motivated maintainer.
link →Ant International, Visa and Mastercard agree a Know Your Agent framework
The three will work on an interoperable KYA framework so merchants and providers can verify which agent is behind a transaction and whether it's authorized, with common methods for linking an agent to a valid entity, scoring its behavior and monitoring activity. Ant's Jiang-Ming Yang put the pitch bluntly: register once, and you don't re-register with the others. There are no technical specs, no named governance body and no implementation timeline — this is three competing protocols agreeing to eventually agree, which is still further than most agent-identity efforts have got.
link →
$ bigger_picture
The trend behind the headlines
Four announcements, four layers, one missing primitive: a verifiable answer to "which agent did this, on whose authority, and what did it touch." The endpoint vendors are solving it with runtime enforcement, the card networks with registration and behavioral scoring, the silicon vendors with hardware-rooted keys for agent-to-agent traffic. None of them are solving it out of foresight — GTIG's six-hour credential campaign is the demand signal, and the industry is retrofitting a control plane onto a deployment wave that shipped without one. Notice who's racing for the position: whoever issues and verifies agent identity ends up holding a toll booth on every agent action that crosses a boundary. If you're building agents now, decide whether you want that identity layer to be something you own, or something you rent from CrowdStrike, Visa, or whoever wins.
Geiger
One read-only command that inventories every agent, harness, MCP server, plugin, hook and AI extension on a machine, and tells you in plain language what each one can reach: filesystem scope, network access, credentials sitting in config files. It's the free local version of the discovery problem CrowdStrike just productized, and the first run on a working dev laptop is usually unpleasant. Run it with npx geiger-scan.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime