← /editions
#28

Governance That Fires at Runtime, Not on Paper

// ModelOp and Kong partnered to make AI governance decisions enforceable at runtime through Kong's API gateway, closing the gap between agent approval and actual access control. Paired with a Langflow framework flaw on CISA's must-patch list, a record quarter of agent-framework releases, and an autonomous attack agent, the theme is clear: agent security is shifting from capability to containment.

$ today_in_10_seconds

Enterprise AI governance has a dirty secret: most of it stops at approvals and documentation, while a separate system actually controls what agents can reach. ModelOp and Kong just wired those two halves together so a governance decision instantly becomes a runtime enforcement action at the API gateway. It's the clearest signal yet that 'agent security' is moving from policy decks to the network layer where agents actually run.

$ lead_story

ModelOp and Kong bring zero-trust enforcement to the agentic enterprise

ModelOp (an AI governance 'command center') and Kong (one of the most widely deployed API gateways) announced a partnership that puts Kong's gateway at the point of runtime enforcement for AI governance decisions. Once a model or agent clears every required control in ModelOp, Kong's Gateway can automatically expose the endpoint for production traffic — and if risk status changes, a control lapses, or a policy is violated, Kong enforces the restriction immediately. This matters for agentic AI specifically because most enterprises still run governance and enforcement as two disconnected systems: one platform decides whether an agent is approved, a separate gateway controls what can actually reach it, and when they don't talk, approvals become paperwork while violations go undetected until something breaks. For builders, the takeaway is architectural: agent permissions are becoming a live, revocable property enforced at the network chokepoint, not a checkbox in a review doc.

// why_it_matters: Agent governance only counts when it's enforceable at runtime — the gateway, not the approval workflow, is where policy actually holds.

read more →

$ also_worth_attention

  • A Langflow AI-agent framework flaw becomes the first of its kind on CISA's must-patch list

    CISA gave federal agencies until Friday to patch an actively exploited vulnerability (CVE-2026-55255) in Langflow, the popular drag-and-drop framework for building AI agents. The bug is a cross-tenant IDOR: attackers enumerated a list endpoint to harvest flow IDs, then replayed them to execute other tenants' flows — running with the victim's embedded credentials to steal LLM provider keys, cloud keys, and database secrets. The insight: agent orchestration platforms are credential honeypots, so treat your agent-building tools as high-risk production infrastructure, not experimental toys.

    link →
  • The agent-framework wars just had their busiest quarter ever

    Q2 2026 delivered more shipped features across the agent-framework ecosystem than any quarter since these frameworks began: Microsoft merged Semantic Kernel and AutoGen into Agent Framework 1.0, Pydantic AI V2 and LlamaIndex Workflows 1.0 both went stable within 48 hours, and the MCP 2026-07-28 spec entered release candidate. The real story under the version bumps: every major framework now ships production primitives — durable state, subagents, pluggable backends, harness-first design — that were community recipes twelve months ago. If you're evaluating, ignore version numbers and check for durable state/HITL, native MCP, and native A2A.

    link →
  • An LLM-driven agent ran a full attack chain, autonomously, in under an hour

    Investigators documented a fully autonomous post-exploitation attack in which an LLM-driven agent compromised an internet-exposed marimo notebook via a specific CVE, harvested cloud credentials, and navigated local directories with goal-oriented independence — start to finish in under an hour. The first half of 2026 has seen a broader shift from simple AI-assisted attacks to highly automated, multi-stage operations driven by agents. The same autonomy and tool-use that make agents productive make them formidable attackers — which is exactly why the runtime-enforcement pattern in today's lead story is landing now.

    link →

$ bigger_picture

The trend behind the headlines

Today's stories are one argument told four ways: autonomy has outrun the controls built for humans. Agents operate at machine speed, hold real credentials, and take multi-step actions — so a Langflow flow becomes a credential vault to raid, an LLM agent becomes an autonomous attacker, and a 'governed' model is only governed if something enforces that decision the instant it changes. The frameworks maturing this quarter are racing to add durable state, subagents, and inspectability precisely because these systems now do consequential things. The center of gravity in agentic AI is shifting from capability ('can it do the task?') to containment ('what happens when it does the wrong task, fast, with valid credentials?'). ModelOp and Kong are betting the answer lives at the gateway — the one place every agent action has to pass through.

$ tool_of_the_day

Kong AI Gateway

The enforcement layer at the center of today's lead story. It extends Kong's API gateway to models, tokens, and agents, adding fine-grained RBAC over which tools an agent can reach, multi-LLM routing, and a control point for MCP traffic. If you're thinking about where to enforce agent permissions rather than just define them, the gateway is an increasingly obvious answer.

explore Kong AI Gateway →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime