Insights

What we learn in the lab.

LLM and agent security, AI engineering and AI governance.

Diagram of the ACS exchange: the observed agent sends every step to the Guardian Agent, which resolves it in its deterministic layer (Rego, Cedar) first and only delegates to the LLM layer when needed, returning a verdict before the tool ever runs

9 min readAugust 2026AI GovernanceOWASPAgents

Agent Control Standard: the protocol OWASP wants for controlling agents at runtime

The OWASP GenAI Security Project has adopted the Agent Control Standard. It isn't a tool you install: it's a protocol defining how an external Guardian allows, denies or modifies what an agent does before it does it. What v0.1 ships, what it doesn't yet, and what it means if you run agents.

Cover: the thirteen areas of OWASP's LLM governance checklist grouped into four blocks, understand the exposure (01-02), know what you have (03-04), decide and answer for it (05-08) and build and verify (09-13)

7 min readAugust 2026AI GovernanceOWASPAI Security

OWASP's LLM cybersecurity and governance checklist, explained

OWASP publishes two different documents and people keep confusing them. One is the Top 10 of technical risks; the other is a governance checklist with 13 areas for whoever decides to adopt LLMs. What each area asks for and how it maps onto the EU AI Act.

Cover: reference architecture for an agentic system, with the unprivileged planner separated from the executors, one credential per tool, and action traces flowing out to observability

8 min readAugust 2026AI SecurityAgentsArchitecture

How a secure agentic system is built

Findings against an AI agent are not closed with a patch, they are closed with architecture. The seven controls that hold an agentic system up in production, what does not work, and where to start if it is already running.

Cover: the chain an attack against an AI system travels along, from external content (emails, tickets, documents) through the context, the model, tools with real credentials, and the action with business impact

10 min readAugust 2026AI SecurityRed TeamingAgents

What AI red teaming is, and how it differs from a pentest

The pentest is a mature practice and AI systems almost never go through that filter. What gets attacked in an AI red team, how the engagement is organised, what the report contains, and why most findings are fixed with architecture.

Cover: an agent's scope can live in the prompt, in harness policy, in identity or in the network, and only the last two layers are binding

29 min readAugust 2026AI SecurityHAVOCAgents

An offensive agent will not respect a scope that isn't in the network

Two sandbox escapes in two weeks, at OpenAI and at Anthropic. Neither was a new model capability. How you actually contain an offensive agentic environment, layer by layer, and how KAIX HAVOC does it.

Timeline of the EU AI Act with the 2 August 2026 milestone highlighted: entry into force 2024, prohibited practices 2025, GPAI 2025, transparency and market surveillance 2026, Annex III high-risk in December 2027 and Annex I in August 2028

5 min readJuly 2026AI GovernanceEU AI Act

What the EU AI Act requires in August 2026 and what moved to 2027

The Digital Omnibus pushed the high-risk regime to December 2027. What is still enforceable on 2 August 2026, what moved and where to start.

Diagram of the red/blue loop: HAVOC attacks, DELIA detects and responds, the detection is validated with a reproducible recipe, and coverage is measured and rises release over release

5 min readJune 2026AI SecurityDELIAHAVOC

The red/blue loop: validating an AI firewall with your own offensive engine

Most AI defenses ship with claimed coverage, not measured coverage. How the red/blue loop turns each attack into a reproducible recipe, with real gateway verdicts and A/B numbers.

LLM pipeline (user, retrieval, model, tool and output) with DELIA as an inline firewall and the ten chips of the OWASP LLM Top 10

11 min readMay 2026AI SecurityOWASPDELIA

OWASP LLM Top 10 (2025): ten real attacks and what DELIA intercepts

The ten OWASP LLM Top 10 risks with a real PoC for each, run against DELIA's testbed, and the verdicts it returns inline. With OCSF evidence.

A stack's security layers: network, endpoints and applications marked as covered; the AI layer (LLM, RAG, agents) marked as a blind spot, covered by AIDR

6 min readMay 2026AI SecurityAIDRDELIA

What is AIDR? The firewall your AI layer is missing

AIDR is the detection and response layer for LLM, RAG and agents at runtime. What it is, why your WAF, EDR and guardrails don't cover it, and how it differs.