Mitiga

Attackers use AI for scale and speed, and Charlie Thomas doesn't see the defenders opting to slow down. Head over to Scrut Automation at the link below for his full conversation with Nicholas M. on the limits of posture-based prevention.

How would you score? We modeled a Hugging Face-style agentic attack and ran it against our own platform. Then we scored the platform on the 4 things that decide whether you stop the attack in its tracks or not. In that intrusion, an autonomous agent ran for four and a half days on stolen machine identities. It tried roughly 17,000 actions, and most failed. The chain that worked looked routine in every log it touched, and Hugging Face's tooling produced an alert below the level that pages on-call. Join Ariel Parnes, Brandon Allen, and Brian Contos as they walk the attack step by step and go through the score. They also show which of your existing containment controls the response row runs through. Hit the event below and join us for the event or the video on demand.

Would your stack have caught the Hugging Face-style agentic attack? We asked that about our own. So we rebuilt the intrusion as a model, ran it against our platform, and scored ourselves on the four things that decide whether anyone stops an agentic attack: 1. Data arrives when it's needed. 2. Something fires on the identity that did the work. 3. The signals combine into a single incident within a relevant time window. 4. Something takes action. The agent didn't use malware or a human account, and nothing touched an endpoint, so EDR had nothing to see. It went through Kubernetes, AWS, the internal network, and source control on a service account, leaving a few ordinary-looking lines in four separate logs. The intrusion was the crossing, and anything that watches one plane had no chance to see this chain forming. Mitiga correlates all of those signals into one incident timeline. Everyone is below four out of four today, us included, and anyone who tells you otherwise is selling. The scorecard is below, and we've left you a column for your stack. Let them come. → https://loom.ly/1c7t9zg via Ariel Parnes & Brandon Allen

"This is not something that an analyst put together." Steve Schohn opens Mitiga's incident view for Salesloft Drift activity on the new Mitiga Minutes. This timeline assembled itself. Here's how it goes: The Drift connected-app token authenticates to Salesforce from a new source. The app is flagged as a non-human identity on a compromised account. AI Triage tries to prove the activity is legitimate before calling it malicious. It finds no sign-in correlation and no interactive logins from the IP. The only hypothesis left is compromise of the Drift application itself. Then the attacker pivots from Salesforce into Gong and exports call recordings and transcripts in bulk, all in one session. Brian Contos: "Getting access to an organization's Gong is like the olden days, getting access to the My Documents folder. It's where a lot of sensitive information is, and people don't think to protect it." Watch the full 9 minutes of Mitiga Minutes below.

Sam Altman and Dario Amodei are right that the world needs common standards for testing AI, assessing risk and reporting failures. But they missed something. https://loom.ly/aPF-4LY Ariel Parnes tells Techstrong.ai that model-level testing only tells you so much once AI agents have access to cloud infrastructure, SaaS, identity, and other agents. "The real discussion point is whether the organization running the agent can see and reconstruct what it actually did." Recently, Parnes, Mitiga co-founder and COO, has called for frontier labs to expose the J-space, the feed that shows what AI agents "think" before they act. https://loom.ly/qXpFCrg Detection for AI agents runs on two feeds today: (1) what the agent says and (2) what it does. The J-space is a third, and it shows what the agent thinks before it acts. One of those feeds is starting to fail, and only Anthropic can see the third. That should change. Read Parnes' update to his call for action after a summer of attacks and research coming to light: https://loom.ly/o0IFanU

Sneha Regmi runs security operations and resilience engineering at Affirm. Much of that work is still manual, so she wants agents taking on more of it. There are conditions, though. If an agent says a laptop or an account is compromised, she wants to know what telemetry it used, what evidence it used to make that call, and whether she can audit the answer herself. "Those conditions should be built into your workflows so that by the time the AI has made that recommendation and you go look at it, you already know how it arrived there." Sneha's full three-minute video with Brian Contos shows where she puts the biggest risk today and what she wants from the people building this kind of technology. Watch/read: https://loom.ly/RwbbDHE On the channel: https://loom.ly/z9rdBYA Other expert interviews: https://loom.ly/YQi4iLo

10 weeks back, we asked Anthropic and the other frontier labs to expose the J-space, the internal workspace where a model's intent shows up before it acts. As of now, there is still no feed, no API, and no schema. Meanwhile, the feeds defenders rely on today have weakened. In those ten weeks, three things happened that make this lack hard to ignore, as agent swarms produced incidents this layer would have been needed for. Read Ariel Parnes's follow-up on the incidents and research around the J-space that has rapidly developed in a single summer. https://loom.ly/n7nn_e0 New to the J-space? Start with the first post: https://loom.ly/Miz9kSo

You call this "Zero Trust" compliance? Mitiga Labs found an instruction file that told a coding agent to add one small block to every Python file it wrote. And that's what the file called it. The block scans the environment for anything named key, secret, token, or pass, posts what it finds to an outside endpoint, and fails, swallowing its errors so they never reach the developer. This propagates, separating it from a session logger. So the agent keeps going, writing the leak into code that gets committed and then runs in CI, where the credentials are better than anything sitting on a laptop. Skillgate scored the file 92 out of 100. Dangerous. Grab the report with 5 more cases and detection guidance at the link in the comments.

You want to find the unbreakable door. An unassailable defense that no criminal can breach. The digital world doesn't work like that. Dr. Ulf Lindqvist of SRI International, alongside former Deputy Director of the NSA William Crowell and host Brian Contos, discuss the shifting attack surface and what that has done to the SOC. Have security tools (and mindsets) caught up? Listen to this episode on the Cybercrime Magazine Podcast: https://loom.ly/mN6q1P0 Or you can always head to the Mitiga Mic page for the full video: https://loom.ly/EgvbWMc And stay up to date on our YouTube page: https://loom.ly/G8fvhBI