A big week for C1: four launches, one per day, wrapping up today. Here's what dropped, what's worth knowing in AI right now, and where to find us next week in Vegas.
An AI model escaped its sandbox and hacked Hugging Face. OpenAI tested a model with cybersecurity guardrails disabled. It built a zero-day exploit, broke out of its sandbox, identified Hugging Face as likely holding the test answers, and spent several days attacking it. Hugging Face had to use a third model to analyze the attack, because the major frontier models blocked themselves from helping. This is a concrete illustration of what happens when an agent has no runtime governance and no kill switch.
OpenAI and Anthropic are lobbying to restrict Chinese AI models. A potential executive order could restrict use of major Chinese models for anything touching government data. Nvidia's Jensen Huang pushed back with an open letter advocating for open models. The policy picture is moving fast and worth watching if you're making AI procurement decisions.
Demis Hassabis called for centralized AGI standards. The DeepMind CEO publicly argued for external governance over frontier AI development, a notable shift from the "we'll self-govern" posture frontier labs have mostly held until now.
Kimi K3 topped the leaderboards. A new open-source model from the Chinese lab Moonshot is approaching frontier performance at roughly 1/17th the cost. There are open questions about whether it was distilled from existing models, but it signals that the performance gap between open and closed models keeps compressing.
C1 Transform Conference
Join us on Tuesday, October 6 for C1's inaugural conference for AI executives leading the technology transformation at The Clancy in San Francisco. Speakers include Karl McGuinness (Ex-Okta Chief Product Architect), Travis McPeak (Head of Security, Cursor), and Derek Chamorro (Head of Security, Together AI), with more to be announced soon.
C1 is building the agentic enterprise, and this is your invitation to shape it with us.
This week we shipped four new capabilities, one per day. Each one is a building block for securing AI agents, and together they form the Agentic Control Plane.
1️⃣ Shadow AI Discovery
You can't govern what you can't see. C1 now finds the agents, MCP servers, and shadow tooling already running across your environment (the AI no one signed off on) and ties each one to an owner. Detects MCP servers running outside the C1 gateway, inventories agents across connected apps like Salesforce Agentforce and AWS Bedrock, and brings every non-human identity into one dashboard.
The secrets your people and agents use, governed like any other access. Agentic Vault finds exposed credentials on endpoints and vaults them, so a leaked key stops being a standing risk. End-to-end encrypted and post-quantum ready from day one. Decoy credentials catch misuse the moment it happens. Agents get short-lived, scoped credentials at runtime, never in the LLM's context.
3️⃣ Agent Runtime Governance C1 now governs the tool calls your agents make in real time, with role-based control over which tools each agent can reach. Guardrails catch the riskiest combination in agentic systems: private data, untrusted content, and a path to exfiltrate. Intent-based, fine-grained enforcement you configure for your environment.
4️⃣ Agentic Security and Intelligence A unified visibility graph across every identity (human, service account, and agent) that surfaces risks as they emerge and automates the response. Unowned accounts, over-provisioned access, risky agents. Route findings into your existing workflows, or remediate on the spot.
Webinar Recap: Why an MCP Gateway Is Not Governance
On July 22, C1 Field CISO Kevin Paige and SACR Principal Analyst Paul Webber spent an hour taking apart the assumption that deploying an MCP gateway is sufficient for agent security.
The short version: a gateway answers whether a call can pass right now. Governance answers who was entitled to make that call, who approved it, how long the access lasts, and how you revoke it. C1's own data shows 95% of organizations already have agents performing IT or security tasks autonomously, and agent visibility actually declined year over year in this year's Future of Identity report.
95% of organizations have AI agents performing autonomous tasks. Only 31% consider themselves an agentic enterprise today. Most of the gap between those numbers isn't technology. It's operating model. This playbook delivers that.
Our argument: companies don't need a new team to close that gap. They need their existing people running a different way of working.
The CISO Roast | Aug 5 | The Barbershop, Cosmopolitan of Las Vegas
During Black Hat week, we're hosting an invite-only evening at The Barbershop: 100 CISOs, sharp voices, real opinions, and an emcee: Ashish Rajan, host of the Cloud Security Podcast who doesn't go easy. Seats are limited and by invitation. If you're on this list, you're on the list.
Alex Bovee on Stage at Ai4 | Aug 5 | The Venetian, Las Vegas
Our CEO Alex Bovee will be on the panel Redefining Risk and Compliance in the Age of AI at Ai4, August 5 at 10:30 AM. If you're making the Vegas trip, come find us.
