Skip to content
A speaker presenting research at a white lectern, with the OAIC 2026 audience in the foreground

OAIC / Public research collection

Talk
Archive

Talks, slides, and ideas advancing the next wave of offensive AI.

Explore the collection

From the conference
to the field.

Many talks are withheld from this archive at the authors’ request due to the sensitivity of the research.

Explore the materials speakers have chosen to make public. Selected 2026 recordings are available now; the full session archive is forthcoming.

12 talksBrowse by topic, speaker, year, or format

Vulnerability research & fuzzing

Agents & assessment

Policy & strategy

Evaluation & model training

Alex Tselevich and Michael Maturi — Offensive Harness Design at OAIC 2026

Offensive Harness Design: Engineering Agents for Static Analysis

Alex Tselevich, Michael Maturi · Mandiant

Mandiant’s Alex Tselevich and Michael Maturi share practical lessons from building agentic harnesses for static analysis, covering threat modeling, lead selection, context and token management, orchestration, deduplication, validation, and benchmarks.

Full presentation and audience Q&A · 1:01:17

Gaspard Baye — Teaching Agents to Escape at OAIC 2026

Teaching Agents to Escape: RL, Command Evasion, and the Limits of the Sandbox

Gaspard Baye, Ph.D. · Valix AI

Gaspard Baye examines how offensive agents search for command-policy bypasses, and why bypassing a filter differs from escaping a capability sandbox. Using SandBench, supervised fine-tuning and reinforcement learning, he compares defenses, analyzes reward-design pitfalls, and explores Bayesian attack graphs as a next step.

Full presentation and audience Q&A · 27:02

Sam Lovejoy — Vibe Fuzzing at OAIC 2026

Vibe Fuzzing: Building an AI-Guided Fuzzing Engine from Scratch

Sam Lovejoy · IBM X-Force Offensive Research

Sam Lovejoy of IBM X-Force Offensive Research explains how VF3 combines graph-based input generation, state-aware fuzzing, and LLM guidance to target code and patches while keeping the fuzzing loop fast. He covers Markov-chain grammars, program state, model context, performance, and cost, followed by a live demonstration and audience Q&A.

Full presentation and audience Q&A · 30:11

No research matches these filters. Try another speaker or topic.

2026 session archive

The full collection of talk videos and slides is forthcoming.
Selected recordings are available in the collection above.

Watch the 2026 keynote

Topic groups are editorial navigation categories. The archive contains currently verified public resources and does not represent the full conference programme.

OAIC Recording in Progress sign for the Three Buddy Problem podcast

Conversations / OAIC 2026

Three Buddy Problem × OAIC

Interviews recorded live at Offensive AI Con with Ryan Naraine and Juan Andres Guerrero-Saade.

Three Buddy Problem on YouTube