$ today_in_10_seconds
Two days ago the EU AI Act's enforcement machinery switched on — and almost everyone got the story wrong. The high-risk regime everyone spent 2025 preparing for has been pushed to 2027, but the penalty regime, GPAI enforcement powers, and the rule that your agent must tell users it's an agent are live now. Meanwhile MCP shipped its biggest spec revision ever, and fresh scans found a third of MCP servers in the wild are exploitable.
$ lead_story
The EU AI Act's enforcement powers went live on 2 August — just not the ones you prepared for
On 2 August 2026, the AI Act's Article 50 transparency and information obligations became generally applicable and enforceable by national authorities across the EU. That means chatbot disclosure, synthetic content marking, and deepfake labelling — and it applies whether or not your system is classified high-risk. The twist: the high-risk obligations did not arrive. Stand-alone Annex III systems now have until 2 December 2027, and AI embedded in regulated products under Annex I has until 2 August 2028, courtesy of the Digital Omnibus on AI. What did land on schedule: the Commission's enforcement powers over general-purpose AI and the full penalty regime — up to 15 million euros or 3% of global turnover. If you ship an agent that talks to EU users, the disclosure duty is a today problem, not a 2027 problem. And if you substantially modify an existing AI system or adapt it for a high-risk purpose, you become the provider in legal terms — technical documentation and conformity assessment shift to you. That's most agent orchestration work, described precisely.
// why_it_matters: Agent autonomy is now a documented, auditable, fineable property of your system — and the clock started two days ago, not next year.
$ also_worth_attention
MCP 2026-07-28 is final — and it's the biggest revision since launch
The headline is that MCP goes stateless at the protocol layer. The initialize/initialized handshake is gone entirely under SEP-2575; protocol version, client info, and capabilities now travel in _meta on every request. The real story is authorization: MCP authz moves from 'technically possible if you wire it up yourself' to following defined RFCs, with a documented path from unauthenticated server to properly secured server behind Okta, Azure AD, or Google Workspace. Resource Indicators bind a token request to a specific MCP server, cutting confused-deputy and replay risk. If you built against 2025-11-25 and assumed sticky sessions, your migration is not cosmetic.
link →A third of MCP servers in the wild are exploitable
BlueRock analysed over 7,000 MCP servers and found 36.7% potentially vulnerable to server-side request forgery; their proof of concept against Microsoft's MarkItDown MCP server pulled AWS IAM access keys, secret keys, and session tokens from an EC2 metadata endpoint. Separately, static and taint analysis across 10,000+ real-world servers found credentials, API keys, and PII leaking at rates above 10%, and the NSA published official MCP hardening guidelines in July. The insight: a single misconfigured MCP server is a cloud credential gateway. Treat every one as an untrusted third party, including your own.
link →TripGain wires MCP into an enterprise T&E API gateway
TripGain launched agentic infrastructure for corporate travel and expense, pairing MCP with its existing API gateway to let agents act across connected travel ecosystems. It's a press release, but the pattern is the one to watch: incumbents with a gateway and real supplier integrations bolting MCP on top rather than rebuilding. The gateway is the moat; MCP is the adapter. Expect every vertical SaaS company with an API to run this playbook by Q4.
link →
$ bigger_picture
The trend behind the headlines
Three stories, one convergence: agents just acquired a compliance surface, a real protocol, and a documented attack surface — in the same week. That's not coincidence, it's the standard maturity sequence for infrastructure, compressed into eighteen months. Note where the deferral came from: the high-risk regime slipped because member states hadn't designated competent authorities and the harmonised standards conformity assessment depends on weren't finished — regulators were demanding conformity against benchmarks that didn't exist. Regulation is now blocked on the same thing your agent deployment is blocked on: nobody has finished building the identity, authorization, and audit layer. You can see the gap in the tooling too — Goose has 52.1k stars and 500+ contributors, and zero SOC 2 or HIPAA certification, which is a hard stop in regulated industries. The teams that win the next year won't have the smartest agents. They'll have the ones whose actions can be attributed to a principal and replayed in an audit.
Goose
An MCP-first open-source agent from Block, now under the Agentic AI Foundation, with 70+ MCP extensions and 30+ LLM providers, shipping CLI, desktop, and API across macOS, Linux, and Windows. Worth running specifically because it forces you to own the authorization boundary yourself instead of inheriting a vendor's — which, after this week, is the muscle you need. Just be honest about the economics: with a top-tier cloud model the API bill lands around $5-20/month for moderate use, so the cost advantage over a flat subscription is modest to nonexistent. The real win is flexibility and privacy.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime