Beyond Guardrails: A Technical Deep Dive into AI Model Safety, Agent Safety & Governance
As AI moves from generating answers to autonomously taking actions, safety changes fundamentally. Model alignment is only one part of the problem. Agents introduce planning, uncertainty, memory, tools, permissions, and entirely new failure modes.
- ✓ Why model safety and agent safety are fundamentally different, and where model-level alignment stops being enough.
- ✓ How uncertainty, long-horizon planning, and partial observability can cause autonomous agents to make unsafe decisions.
- ✓ How to architect safety when the intelligence itself is probabilistic and non-deterministic.
- ✓ How to evaluate, govern, and establish trust in increasingly autonomous AI systems.
Register for the session
Thursday, September 10, 2026. 9:00 PM IST / 11:30 AM ET / 7:30 PM GST. The joining link is sent to your email once you register.
Who you will be hearing from
Arnab Chattopadhayay
More than 23 years solving hard IT security problems, with leadership roles at British Telecom, Tech Mahindra, iViZ (Synopsys), MetricStream, Capgemini, and IBM. Deep grounding in the engineering behind autonomous offensive security, including the guardrails, cost controls, and oversight models that decide whether an agent is safe to run inside a live enterprise.
Other Turbo Talks sessions
Register for any session in the series. Each one is a standalone 20-minute live talk.
AI is collapsing the attacker timeline, exposing "depth debt" in current ASM and CTEM tools. Why the gap compounds, and how to close it.
Register
Black Hat shows where the industry is going. DEF CON shows how attackers plan to break it. Field notes on what shifted this year.
Register90 days on live bug bounty programs. 204 valid reports. Top 3 across multiple HackerOne leaderboards, using AI agents alone.
Register
Two decades of secure coding, SAST, DAST, and Shift Left, and the same vulnerability classes keep reaching production.
RegisterAs offensive AI agents get more capable, control becomes the harder problem. Practical guardrails for deploying them safely.
Register
Every vendor claims AI now. The questions that separate AI-native platforms from scanners with an AI wrapper.
Register