← /editions
#57

Amazon Lost. Your Agent Can Shop.

// A federal appeals court ruled that when an AI agent acts on user direction, the user is the one accessing third-party servers — the first appellate word on agent liability under the CFAA. Plus: the EU quietly moved its high-risk AI deadline, Cyera paid $1B for non-human identity, and MCP shipped its biggest spec revision since launch.

$ today_in_10_seconds

A federal appeals court just ruled that when your agent buys something, you did the accessing — not the company that built the agent. It's the first appellate word on agent liability under the CFAA, and it landed the same week the EU quietly gutted its own high-risk AI deadline and a data-security firm paid $1B for the ability to tell one agent from another. Capability wasn't the story this week. Control was.

$ lead_story

Ninth Circuit vacates Amazon's injunction against Perplexity's Comet agent

On August 4, the Ninth Circuit vacated the preliminary injunction that had barred Perplexity's agent from reaching Amazon.com on users' behalf, finding Amazon unlikely to win on its Computer Fraud and Abuse Act and CDAFA claims. The panel, authored by Judge Milan D. Smith Jr. in No. 26-1444, held that Comet's Assistant is a tool operated by users rather than by Perplexity, applied the rule of lenity to resolve CFAA ambiguity against liability, and noted there is almost no caselaw on assigning responsibility for AI agents under the statute. The court's reasoning was mechanical rather than philosophical: Comet receives screenshots that the user's own browser captured. Amazon's trademark and state-law claims survive, the case returns to district court, and the panel was explicit that the holding is narrow and that agentic AI law will change.

// why_it_matters: Liability for what an agent does sits with the human who directed it, not the vendor who shipped it — which is either the best news your legal team got this quarter or the scariest.

read more →

$ also_worth_attention

  • The EU AI Act deadline everyone circled was already moved

    The Digital Omnibus on AI — Regulation (EU) 2026/1744 — entered into force on July 27, deferring standalone Annex III high-risk obligations to December 2, 2027 and embedded Annex I systems to August 2, 2028. What actually became enforceable on August 2 is narrower: Article 50 transparency duties and the Commission's enforcement powers over general-purpose models, with fines up to 15 million euros or 3% of worldwide turnover. Systems already on the market get until December 2, 2026 for machine-readable watermarking. If your compliance roadmap still says August 2026 equals conformity assessment, it's describing a law that no longer exists.

    link →
  • Cyera pays ~$1B for non-human identity

    Cyera — fresh off a $600M raise at a $12B valuation — signed an LOI to acquire Oasis Security for roughly $1B, mostly cash, to monitor agent behavior and govern what agents are permitted to reach. Cyera's own number for why: non-human identities inside Fortune 500 companies grew nearly 500% in six months and are now the fastest-growing identity type in the enterprise. For calibration, SailPoint bought Entro Security six weeks earlier for a reported $200M — Oasis was a bigger business, but not five times bigger, so a chunk of that premium is the agentic narrative itself.

    link →
  • MCP's biggest spec revision since launch shipped July 28

    The 2026-07-28 specification makes MCP stateless at the protocol layer, and it contains breaking changes — implementers have work to do. Authorization is the part worth reading twice: MCP servers are now formally OAuth 2.1 resource servers, aligned with OpenID Connect, plus a new extensions framework with reverse-DNS identifiers and independent versioning. The practical migration test is whether your requests can move across server instances without losing context — if they can't, you have a hidden session dependency to find.

    link →

$ bigger_picture

The trend behind the headlines

Three questions have been hanging over agentic deployment for two years: who's responsible when an agent acts, who's allowed to let it act, and how does the other side of the connection know what it's talking to. This week produced a partial answer to each — narrowly, from a court; loosely, from Brussels; expensively, from the M&A market. Notice that none of them was answered by a model release. The Ninth Circuit's reasoning is also the most fragile of the three: it rests on Comet needing user direction, which means the more autonomous your agent gets, the less that logic protects you. Build as though the user-as-actor framing has a shelf life, because the court more or less said it does.

$ tool_of_the_day

Web Bot Auth

The missing half of the Perplexity ruling. It's an IETF architecture that lets agents cryptographically identify themselves using HTTP Message Signatures, replacing IP allowlists, User-Agent strings, and shared API keys with a per-request signature validated by the origin. Built on RFC 9421 with Ed25519 keys, a Signature-Agent header, and a published JWKS directory, backed by Cloudflare, Amazon, Akamai, and OpenAI, with an IETF working group chartered in 2026. If you operate an API or a storefront, this is how you say yes to some agents and no to others without guessing.

explore Web Bot Auth →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime