DSOLG November Event
About the event
Details
Welcome to the DevSecOps London Gathering November Event on Wednesday 18 November! We bring you two more great speakers, as well as the usual conversations, pizza and beer!
📍 Hosted at Adaptavist, Telephone House, 69-77 Paul Street, London, EC2A 4NQ 📅 Wednesday, 18 November 2026 🕕 6:00–8:00 PM
Talk Abstracts:
Talk 1 - Nobody inspects hop three — securing agent-to-agent messages Multi-agent AI pipelines split work across orchestrator, researcher, and writer agents — but most teams still only inspect the user prompt. Indirect injection arrives inside web fetches and inter-agent handoffs, and the last agent in the chain often holds the credentials. This session tells that story live: an attack that succeeds without protection, then the same pipeline secured with AgentGuard — open-source Python middleware (`pip install inter-agent-guard`) that inspects every hop with rules, optional ML scoring, Ed25519 trust attestation, YAML capability manifests, and a tamper-evident audit log. You will see which defences are deterministic and deployable today (signatures and policy) versus probabilistic (ML with published limits). The demo uses IntelBrief, a three-agent LangGraph pipeline: poisoned web content and forged orchestrator messages blocked at the tool boundary and signature layer. Audience takeaway: deploy identity and capability controls immediately; treat ML as defence in depth with honest metrics; audit every block decision.
Speaker Bio: Umme Nizba designed and built AgentGuard, an open-source firewall for the messages AI agents send each other — inspecting content, verifying Ed25519 signatures, enforcing YAML capability policy, and logging every decision in a tamper-evident audit trail. The library ships on PyPI and GitHub (v1.2.0) with IntelBrief, a live three-agent demo for indirect injection and impersonation attacks.
They hold an MSc in Information Security from Royal Holloway University of London and work as a software engineer in regulated financial services, with experience spanning secure system design, threat modelling, PCI-aware data pipelines at scale, authentication, and real-time platforms. Their talks emphasise what teams can deploy deterministically tomorrow — trust and policy — versus probabilistic ML with published limits.
