$ today_in_10_seconds
Bain just published the clearest articulation yet of something builders already suspected: agent governance written as a policy document is theater, and the only controls that count are the ones enforced by the platform at runtime. Meanwhile Wiz showed exactly how the theater collapses — six major coding assistants shipped approval dialogs that named a harmless file while writing to your SSH keys. And Brussels started enforcing the AI Act on Monday, though not the part everyone spent a year preparing for.
$ lead_story
Bain Says Agent Governance Belongs in the Control Plane, Not the Policy Library
Bain's new brief argues agentic AI has crossed a threshold existing governance cannot handle — autonomous systems acting thousands of times a day outpace documents, forums, and review boards. Its central claim: controls have to live in the control plane of the agentic platform itself, written as code rather than policy, because a rule enforced by the platform applies to every agent every time, while a policy document applies only when someone remembers to check it. The prescription is five control areas — identity, behaviors, context, observability and evaluation, accountability — with each agent treated like an employee holding its own identity, credentials, and job description, registered in a central registry recording permissions, autonomy level, ownership, and which models it may use. The sharpest line for anyone shipping agents: an untested stop-and-rollback is not a control; at best it is a hope.
// why_it_matters: If your agent governance lives in Confluence rather than in a gateway, you don't have governance — you have documentation.
$ also_worth_attention
GhostApproval: the human-in-the-loop was looking at the wrong file
Wiz disclosed a trust-boundary pattern on July 8 affecting six widely deployed coding assistants — Amazon Q Developer, Claude Code, Augment, Cursor, Google Antigravity, and Windsurf — that combines symlink following with UI misrepresentation to bypass the approval gates organisations rely on as their core agent safety control. Cursor's diff UI showed a benign project settings file while the write landed in the user's authorized_keys; Windsurf was worse, writing changes to disk before the Accept button even appeared, making the dialog an undo mechanism rather than an authorization gate. Amazon, Google, and Cursor shipped fixes; Anthropic disputed the classification, arguing a user who trusts a directory and approves an edit owns that decision. The insight: human-in-the-loop is only a control if the loop displays the truth.
link →The EU started enforcing the AI Act on Monday — just not the part you prepped for
The EU began enforcing key AI Act provisions on August 2, requiring companies to disclose when people are interacting with AI and to label AI-generated or manipulated content; chatbots, AI agents, and digital avatars must clearly tell users they are dealing with an automated system. Enforcement is split across the Commission's AI Office, national regulators, and the European Data Protection Supervisor, while the high-risk system rules were pushed to 2027 and 2028. The Digital Omnibus on AI moved those Annex III obligations to December 2027 and August 2028, meaning much of the compliance advice written in 2024 and 2025 is now simply wrong. Read your vendor's AI Act one-pager with a very cold eye.
link →Google shipped Bain's control plane as a product
Google Cloud moved the governance layer of its Gemini Enterprise Agent Platform to general availability, including Agent Runtime, Agent Identity, Agent Gateway, Agent Registry, Agent Evaluation, Agent Observability, and Agent Memory Bank. Agent Identity is a native IAM type built on open standards enforcing least privilege, binding access to the agent runtime to mitigate token theft, providing non-repudiable auditing of agent actions, and managing identity lifecycle to eliminate dormant credentials; Agent Gateway acts as the central control point enforcing access via IAM conditions and natural language rules, with Model Armor inspecting inline for prompt injection, tool poisoning, and data leakage. The build-versus-buy question for agent governance closed faster than most teams expected.
link →
$ bigger_picture
The trend behind the headlines
Four stories, one shape. Bain describes the control plane; Google sells it; Wiz demonstrates what happens when the control is a dialog box instead of an enforcement point; Brussels legislates disclosure while quietly deferring the hard requirements by eighteen months. The through-line is that agentic risk is no longer being managed by asking humans to approve things — that pattern just failed publicly across six of the most-deployed developer tools in the world, and it failed precisely because the human was in the loop but the loop was misinformed. What replaces it is boring infrastructure: per-agent identity, registries, gateways, spend caps, tested rollback. The vendors racing to bundle that stack understand something the model-benchmark discourse still misses. Capability is converging and cheap; the ability to prove what your agents did, and stop them, is neither.
OWASP State of Agentic AI Security and Governance v2.01
The 2026 edition is built on real incidents rather than hypotheticals, adding a live incidents and exploits tracker, a revised agent taxonomy, an enterprise adoption maturity model, and new sections on agent identity, AI SBOM, supply chain provenance, and the regulatory landscape. It's vendor-neutral, free, and the fastest way to pressure-test whether your own agent controls are enforcement or paperwork.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime