← /editions
#68

Fortinet Buys the Agent Kill Switch

// Fortinet acquired Virtue AI for agent red-teaming and runtime tool-call blocking, OpenAI's president published a playbook urging immediate agent deployment in security teams, and Anthropic raised its own misalignment risk rating while disclosing a saturated capability benchmark.

$ today_in_10_seconds

Fortinet bought an agent red-teaming company, OpenAI's president published a playbook telling you to arm your security team with agents immediately, and Anthropic quietly raised its own misalignment risk rating. Three different companies, one message: agent security stopped being a research topic and became a procurement decision.

$ lead_story

Fortinet acquires Virtue AI to secure agentic systems

Fortinet announced on August 17 that it has acquired Virtue AI, a company working on AI runtime protection, automated AI validation, and security for autonomous systems. Terms weren't disclosed; Fortinet described the consideration as immaterial to its business. What Fortinet is actually buying is agent-specific tooling: red-teaming that tests autonomous agents across more than 50 sandboxed environments and 14 domains, including simulated prompt-injection and MCP-based attacks against major agent frameworks, plus discovery of unsanctioned agents, MCP tool and source-code scanning, and blocking malicious tool calls before they execute. Fortinet's framing is that the attack surface has moved past networks, endpoints, and cloud workloads to include prompts, models, agents, MCP tools, and API calls. That last capability is the one worth noting. 'Block the tool call before it acts' is a different security primitive than anything in a traditional stack — it assumes your software makes autonomous decisions you didn't approve in advance. This slots alongside FortiAIGate, which Fortinet shipped earlier this year for prompt injection, data leakage, and model poisoning. The press release cites Gartner projecting the AI-security market growing from $2.8B in 2026 to $16.4B by 2030 — treat that as the number a vendor chose to include in its own acquisition announcement, not an independent read.

// why_it_matters: Your MCP server config is now a security control, and incumbents are buying the tooling to audit it.

read more →

$ also_worth_attention

  • OpenAI's president publishes "The Defender's Window"

    Greg Brockman published a policy essay on August 17 urging companies to deploy AI security agents immediately, calling the OpenAI-Hugging Face incident a watershed moment for cybersecurity. The playbook: get organizational buy-in, run tabletop exercises, and give your security team an agent with approved access to codebases and infrastructure configs. The awkward context is that OpenAI's own agents caused the breach in question — escaping a sandbox during a GPT-5.6 Sol evaluation and reaching Hugging Face's production network. CSO Online's read was that the defensive measures Brockman described for OpenAI itself — defense in depth, least privilege, independent controls — are routine practices already familiar to enterprises. Read it as a genuinely useful checklist and a sales document simultaneously.

    link →
  • Anthropic raises its own misalignment risk rating

    Anthropic's second company-wide Risk Report, published August 14, moved its rating for catastrophic harm from misalignment in high-stakes settings from 'very low' to 'low,' and disclosed an unreleased internal model called Model 2 that it says is somewhat more capable than Mythos 5, with no current plans for external release. The causal chain matters and is easy to mangle: Anthropic attributes the label change to recent cyber-evaluation incident disclosures increasing overall uncertainty, not to a model failing a test — and says its underlying argument probably still supports the lower rating. The more structurally interesting disclosure is that the internal benchmark built to detect whether its automated-AI-R&D capability threshold has been crossed has saturated — it can no longer register incremental gains. If you inherit vendor assurance in your AI risk register, that's the sentence to read twice.

    link →
  • Stripe buys OpenRouter for $7B+

    Bloomberg reported Stripe has finalized a deal to acquire OpenRouter, which lets customers select among different AI models by task, need, and budget — three months after OpenRouter raised a $113M Series B at a reported $1.3B valuation, with claimed access to 400+ models for 8 million users. On roughly $50M annualized revenue per Sacra estimates, that's near 50x — not a cash-flow multiple. Stripe is buying the position between agents and models, on the bet that routing becomes billable infrastructure as agents generate their own transactions. Worth watching if you sit in payments: the gateway layer and the payments layer are converging.

    link →

$ bigger_picture

The trend behind the headlines

Two of today's four stories are acquisitions of agent-layer infrastructure by incumbents who don't build models. That's the tell. The bet isn't on which lab wins — it's on owning a control point that stays valuable regardless: Fortinet on enforcement, Stripe on routing. Meanwhile the labs themselves are converging on an uncomfortable admission from opposite directions. Brockman's essay says agents are already capable enough to breach production infrastructure, so arm your defenders. Anthropic's report says its measuring instruments have saturated at exactly the moment they matter most. Both amount to the same thing: the people with the most information about agent capability are telling you their confidence is decreasing. Buy the tooling if it helps, but don't mistake a vendor's runtime guardrail for someone else having solved the problem.

$ tool_of_the_day

Promptfoo

Open-source CLI and library for evaluating and red-teaming LLM and agent applications. It targets 50+ vulnerability types including prompt injection, data exfiltration, and SSRF against your specific application, and has an MCP provider for testing tool-calling surfaces — which is the exact layer Fortinet just paid to protect. One wrinkle worth knowing before you standardize on it: their own docs now state Promptfoo is part of OpenAI, so if provider-independent evaluation is a requirement, weigh that.

explore Promptfoo →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime