Nexus Lab (the joint research laboratory of INT2 and VNU Information Technology Institute) attended Security Bootcamp 2026, held on 10–12 September 2026 at Hai Ba Trung Hotel, Buon Ma Thuot, Dak Lak. Organised by the Vietnam Internet Association (VIA), the thirteenth edition of the bootcamp ran under the theme Agentic Security: AI that not only assists, but actively defends and actively attacks.
The three-day programme brought together the national security community with 27 talks selected from submissions across research institutions, security vendors and enterprise teams. Nexus Lab attended as a research participant, following both tracks most relevant to our work: how AI is applied to defence, and how AI systems themselves become attack surface.
What the programme covered
Reading across the talks, four recurring threads stood out:
- AI both attacks and defends – and the asymmetry is widening. Multiple speakers showed attack costs dropping toward zero with AI assistance while defensive costs remain flat. The defensive answer proposed across talks was not more tooling but structural change: threat modelling at design time, security champions embedded in development teams, and permission models that assume AI agents will act on untrusted content.
- Prompt injection is a systems problem, not a filter problem. A recurring conclusion: filters and detectors catch known injection patterns but cannot give an architectural guarantee that untrusted data will not become instruction. Proposals ranged from separating instruction channels inside model architectures to deployment guardrails – bounded delegation, tool boundaries with two-man rules, egress checks and audit trails tied to context hashes.
- Agentic SOC operations are moving from assistance to governed autonomy. Teams running agentic triage in production reported alert triage dropping from 25–30 minutes to roughly two minutes with human-in-the-loop controls. The shared principle: autonomy expands and contracts based on evidence, not on model capability alone.
- AI-accelerated offensive research is now routine work. Teams demonstrated unauthenticated RCE chains on AI gateways and IoT firmware, BAC zero-days found by observing authorisation sinks at runtime, and supply-chain attacks that target model, skill and MCP trust boundaries. The pattern is consistent: AI compresses the labour between hypothesis and proof-of-concept, while the researcher’s judgement on scoping, target selection and verification remains the bottleneck and the safeguard.
Nexus Lab’s post-event report, shared with the community
As a research unit, Nexus Lab has prepared a summary report of the talk content, built in a disciplined, objective and machine-assisted way:
- The 25 publicly released slide decks were collected unchanged from the organiser’s official repository and their text extracted with machine readers (PyMuPDF, python-pptx) – no manual retyping, every figure traceable to its source slide.
- Quantitative claims, CVE identifiers and figures are reproduced as printed on the slides, with the source the slide itself cites; the report does not rank or grade talks and adds no interpretation beyond the source material.
- The talks are grouped into six themes by the speakers’ own declared topics rather than editorial judgement, with a short three-sentence summary per talk: content, problem addressed, result.
The short version of the report (v1.1, covering 26 of 27 talks – one talk on Apple Secure Kernel remains unpublished work-in-progress research) is available as a PDF:
Security Bootcamp 2026 – Báo cáo nội dung các bài trình bày (bản tóm tắt, v1.1, PDF)
The full version is available on request – please contact us. The complete slide decks belong to the organisers and speakers; please reach out to the Security Bootcamp 2026 organisers directly for access.
We thank the Vietnam Internet Association and the programme committee for the edition in Buon Ma Thuot, and the speakers for publishing their material.