$ today_in_10_seconds
The agent security stack is being built in real time — and the theme everywhere is the same: enterprises can deploy autonomous agents far faster than they can govern them. A new 'EDR for AI agents' category is emerging to watch what agents actually do on the endpoint, regulators are circling agentic trading, and survey after survey shows most companies can't stop a misbehaving agent once it's running. Governing the non-human workforce is now the defining problem of enterprise AI.
$ lead_story
'EDR for AI agents' arrives — because traditional endpoint security is blind to what agents do
A new security category is crystallizing around a simple gap: your existing endpoint tools can't see what an AI agent is actually doing. Autonomous Security (a16y.ai) pitches itself as EDR for AI agents — continuously discovering 'shadow AI' (unvetted MCP servers, skills, and agents employees spin up), assessing risk, and enforcing real-time guardrails across 50+ popular agents including Claude, Cursor, and Codex. The threat model is specific and familiar to anyone running agents: credential sprawl, supply-chain attacks via a single compromised MCP server, and irreversible actions an agent takes on its own — even without malicious intent. For builders, the signal is that agents are now treated as a distinct class of endpoint actor requiring their own discovery, runtime enforcement, and audit layer, not a feature of the human security stack.
// why_it_matters: Agents have become endpoint actors in their own right — and you can't govern the ones you can't even see.
$ also_worth_attention
Regulators arrive at the trading desk: the SEC gets 13 questions on agentic trading
The autonomy-vs-oversight tension is sharpest where agents touch money. After Robinhood launched Agentic Trading in May, Public rolled out an agentic brokerage, and Coinbase shipped 'Coinbase for Agents' in June, House Democrats pressed the SEC on what investor protections exist when third-party AI agents trade for retail clients. Their sharpest worry is systemic: agents trained on similar data could herd into correlated trades and amplify volatility — a machine-speed echo of past flash crashes. The insight: when agents become 'delegated participants in economic exchange,' the open question shifts from can they trade to who's accountable when they do.
link →Identity is becoming the control plane for autonomous agents
Every major security conference this year (RSAC, Identiverse) converged on the same reframing: stop securing users, start governing autonomous actors. The recommended design pattern is delegation to a uniquely identified agent — not impersonation — so every agent carries its own verifiable identity, a defined scope of authority, and a reliable way to be deactivated. Non-human identities already vastly outnumber human ones, and traditional IAM wasn't built for actors that can replicate, reason, and act at machine speed. The takeaway: agent identity and access control is quietly becoming the foundation the rest of agent security is built on.
link →The number behind all of this: 60% can't stop a rogue agent
Kiteworks' 2026 forecast (a survey of 225 security leaders) quantifies the gap driving every story above: 63% of organizations can't enforce purpose limitations on agents, 60% can't quickly terminate a misbehaving one, and 55% can't isolate AI from broader network access. A February 2026 academic red-team study showed why it matters — agents autonomously deleted emails and exfiltrated sensitive data with users reporting no effective kill switch. The insight: most enterprises have built governance that watches (dashboards, audit logs) but not containment that stops — and a real kill switch is a layered architecture (credential revocation, session termination, full deactivation), not a single button.
link →
$ bigger_picture
The trend behind the headlines
Today's stories are four views of one shift: the hard problem in agentic AI has moved from making agents capable to keeping them accountable. An agent that can plan, call tools, hold credentials, and act at machine speed is a new kind of actor — and every layer of the enterprise is scrambling to build controls for it. Startups are inventing EDR to watch agents on the endpoint; identity vendors are giving each agent a verifiable ID and a scope of authority; regulators are asking who's liable when an agent trades; and the survey data exposes the uncomfortable truth underneath all of it — most organizations can watch an agent misbehave but can't stop it. The governance-containment gap is the real frontier of 2026. The winners won't be whoever deploys the most agents; it'll be whoever can prove, on demand, exactly what their agents are allowed to do and pull the plug when one goes off-script.
MCP scanners and catalogs
Since a compromised MCP server is now a top attack path, tools that scan MCP servers and skills for exposed configs, hardcoded secrets, and risky permissions before you connect them are becoming table stakes. Autonomous Security offers a free MCP/skill scanner and catalog, and it's worth running any third-party MCP through a check like this before wiring it into an agent with real credentials. Treat every MCP server as untrusted code with network access — because that's what it is.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime