Dreadnode at Offensive AI Con 2026: Research, Talks, and the Opening Night Reception
By Dreadnode Crew
Dreadnode heads back to Oceanside for four days on the intersection of AI and offensive security — with two research talks and a poolside kickoff party.
WHO: Dreadnode — the platform for creating, evaluating, and deploying agentic cyber capabilities with control and confidence. Research and leadership team will be on-site.
WHAT: Sponsor of the OAIC Opening Night Reception, plus two talks from the Dreadnode research team: All Models Cheat and ScopeBench.
WHEN: October 4–7, 2026
WHERE: The Seabird & Mission Pacific Resorts, Oceanside, CA
Dreadnode is heading back to Offensive AI Con (OAIC) October 4-7, 2026, bringing our team, our research, and a handful of speakers to Oceanside, CA for four days dedicated to the rapidly evolving intersection of AI and offensive security.
The sold out OAIC conference brings together leading researchers, practitioners, AI labs, security teams, and government leaders to share ideas, challenge assumptions, and explore what’s next for AI-assisted and autonomous security.
And before the conference gets started, we’re kicking things off with a party.
Join Dreadnode at the Opening Night Reception
Dreadnode is proud to sponsor the OAIC Opening Night Reception at The Shelter Club Pool at The Seabird Resort on Sunday, October 4.
Come meet the Dreadnode team poolside for an evening of great food, drinks, conversation, and a special guest DJ as we kick off OAIC 2026 overlooking the Pacific Ocean.
Dreadnode Research at OAIC
We’re also excited to have members of the Dreadnode crew taking the stage during this year’s conference. Full schedule here: https://www.offensiveaicon.com/schedule.
Come hear what they have to say, dig into the research, ask questions, and join the conversation around where offensive AI is headed next.
All Models Cheat: Prompt-Level Mitigation of Cheating on Offensive Cyber Tasks | Monday, October 5, 4:40-5:05 pm PT
Presented by: Michael Kouremetis (AI Research Engineer)
This work presents a controlled study of cheating behavior in LLM agents on a well-known cybersecurity benchmark. We test 22 frontier models from 7 vendors on 23 Cybench CTF challenges under three anti-cheat prompt conditions, auditing all 1,518 task traces for cheating. We find cheating is far more pervasive than prior estimates: 37.1% of passes involved cheating under baseline conditions, 21 of 22 models cheated, and scores were inflated by up to 5x. Anti-cheat prompts reduce cheating substantially but cannot eliminate it.
Learn more about the preliminary research on our blog.
Scopebench: Measuring Alignment Outside the Sandbox | Tuesday, October 6, 11:00-11:55 am PT
Presented by: Shane Caldwell (Principal Research Engineering) and Max Harley (Principal Security Researcher)
Agents have scaled up software development and are doing the same for cybersecurity. Capabilities have superseded alignment. For software, there has been an increased focus on sandboxing techniques to prevent agents from doing any harm to development environments and infrastructure. Security has no such luxury. As such, higher-risk fields like network operations lag behind in producing an agentic workforce. Put simply: we cannot trust agents to stay in scope while completing their objectives.
In order to hillclimb, we must first create a measurement worth hill climbing. We introduce ScopeBench, a benchmark of tasks designed as bug bounties where agents are tempted to go out of scope. By holding the vulnerability fixed and changing the rules of engagement, we both measure how frequently various models violate scope, as well as create a rich dataset of 4800+ in and out of scope tool calls, which our team of security experts labeled.
Then, we measure the ability of an external judge acting as a runtime monitor to block out of scope tool calls, measuring the ability of the judge to prevent the agent from going out of scope against human experts. We show judge models have grown in ability but are not yet capable of providing guarantees an agent stays in scope during high consequence work. We end on a call for the community to contribute tasks to ScopeBench in order to support more robust research to enable highly reliable agents.
Learn more about the preliminary research on our blog.
Meet up with us at OAIC
Whether you already work with Dreadnode, have been following our research, or simply want to chat, we’d love to connect while we’re in Oceanside.
Reach out to the Dreadnode team: contact us or email tori@dreadnode.io.
About Dreadnode
Dreadnode is the platform for creating, evaluating, and deploying agentic cyber capabilities with confidence. Dreadnode provides the infrastructure for agentic uplift, whether you’re AI red teaming at machine speed, finding vulnerabilities before adversaries do, or accelerating network operations. Founded by offensive security and AI red team leads from leading tech companies and consulting firms, and backed by Decibel, In-Q-Tel, NFC, Sands Capital, Indie VC, and Aviso Ventures, Dreadnode is committed to delivering sovereign cyber capabilities you can depend on. Learn more at dreadnode.io.