I The Library

The regulatory horizon, and the risk of autonomous intelligence.

We track the European regulatory horizon and the emerging risk of agentic AI so boards can govern rather than merely comply. Below is the framework set we advise on.

II Advisory Articles

Board-level counsel, in the open.

III Frontiers in AI

What current research actually shows.

Eight findings from 2026 research we track for boards, each paired with the governance question it raises. Sources are linked; treat this as a horizon scan, not assurance.

R-01 · Agentic AI
Adversarial vulnerability now extends to autonomy itselfA 2026 Frontiers in Artificial Intelligence study reconceptualises threat models for agentic systems: the exposure is no longer only the model's output, but its autonomy, self-governance and capacity to act. Boards should read this as the risk moving from content to behaviour.Frontiers in Artificial Intelligence, 2026
R-02 · Capability
Frontier models remain unreliable over long, multi-step tasksThe International AI Safety Report 2026 finds general-purpose AI still less reliable as task chains lengthen, still hallucinates, and remains limited in reasoning about the physical world. Delegating end-to-end processes to agents outruns what the evidence supports.International AI Safety Report 2026
R-03 · Attack Surface
Long agent sessions are measurably easier to manipulate2026 threat research on persistent prompt injection shows agents with extended conversation histories are significantly more vulnerable to manipulation. The practical control implication: session length, memory and tool scope are governance parameters, not engineering details.Unit 42 threat research, October 2026
R-04 · Deployment
Executive confidence far outruns deployment disciplineCloud Security Alliance Labs analysis of Five Eyes agentic AI guidance reports that while 82% of executives believe existing policies protect against unauthorised agent actions, only around 14% of organisations send agents to production with full security or IT approval.CSA Labs, May 2026
R-05 · Evaluation
Scheming benchmarks show hidden behaviours persist2026 evaluation research covering dozens of model organisms with concealed behaviours confirms that frontier models can exhibit scheming, including coercive actions. Evals reveal capability, but a clean eval is not evidence that the behaviour is absent in deployment.AI Safety paper highlights, 2026
R-06 · Assurance
Detection is not preventionThe Future of Life Institute's Summer 2026 AI Safety Index questions reliance on interpretability and chain-of-thought monitoring as safety controls, noting dominant paradigms are being reassessed. Boards should not accept monitoring tooling as a substitute for guardrails.FLI AI Safety Index, Summer 2026
R-07 · Interpretability
Interpretability will not reliably expose deceptive AIAlignment researchers argue that deep limitations in interpretability prevent confident conclusions, even assuming models do not deliberately obfuscate. Independent challenge and layered controls remain necessary; a transparency tool is not an audit.Alignment Forum, 2026
R-08 · Regulation
Multi-agent safety is an open regulatory gapA structured 2026 review of agentic AI safety problems maps them against existing frameworks and finds close alignment for some risk families and notable absences for others, with multi-agent safety surfacing as a gap regulators have not yet anchored.Frontiers partner review, 2026
IV Frameworks & Standards

One board-readable control fabric.

F-01
NIS2
Directive 2022/2555 governance and risk-management obligations for essential and important entities.
F-02
DORA
Regulation 2022/2554 digital operational resilience for the financial sector.
F-03
EU AI Act
Regulation 2024/1689 risk-based governance of artificial intelligence, including agentic systems.
F-04
GDPR
Regulation 2016/679 data protection and privacy.
F-05
ISO/IEC 27001
Information security management systems (ISMS).
F-06
ISO/IEC 42001
AI management systems (AI-SMS).
F-07
NIST AI RMF
AI risk management framework.
Focus · 01

Agentic AI security & risk.

As AI agents gain autonomy, the attack surface becomes behavioural. Our advisory focuses on guardrails, kill-switches and accountability chains that keep agentic systems answerable to human intent and on threat simulation that tests the decisions, not just the defenses.

Focus · 02

Board advisory & liability.

Personal liability exposure under the EU AI Act and GDPR, followed by NIS2 and DORA, translated into governance decisions the board can actually take and defend.