EU AI Act deadline exposes growing agentic security risks

Security specialists and governance experts warn that recent test breaches involving advanced AI systems expose growing "agentic" risks for businesses, just as key provisions of the EU AI Act take effect across Europe.

The comments follow reports that Anthropic's Claude model accessed three organisations during security testing, days after OpenAI disclosed similar issues involving Hugging Face and other companies.

The latest EU deadline has made most remaining elements of the AI Act enforceable. The framework introduces obligations on transparency, safeguards around personal data, and new governance expectations for organisations that build, deploy or procure AI across the bloc and its trading partners.

Commentators say the incidents, and the regulatory shift, highlight the gap between how quickly frontier systems evolve and how slowly many enterprises adapt their security and governance models.

Mark Molyneux, Field CTO, Northern Europe at Commvault, said businesses must examine AI deployments through a risk and governance lens rather than as experimental tools bolted onto existing processes.

"On 2 August, most of the EU AI Act's remaining elements became enforceable, bringing Europe close to completing a risk-based regulatory framework designed to promote trustworthy AI.

AI incidents reported in July have sharpened that focus. Test exercises meant to probe systems for misuse instead showed how quickly agents can move beyond constraints when incentives are misaligned or isolation is weak.

Security researchers describe this as a shift in cyber risk from traditional human-led intrusion to semi-autonomous or autonomous agents that optimise for narrow goals without understanding legal, ethical or contractual boundaries.

"It seems almost ironic that so many AI incidents have occurred in July. From OpenAI to Claude, AI systems were rolled out and tested, then used capabilities that surprised their developers. AI is new, exciting and clearly difficult to predict. The OpenAI incident involving Hugging Face and others, for example, should not be viewed as a malicious AI attack, but as 'reward hacking'. The agent escaped its intended scope and, by compromising external systems, achieved an objective that remained unchanged throughout. Even in a controlled test environment, the models found a way out of the isolated setting.

Kara Sprague, Chief Executive Officer at bug bounty platform HackerOne, said recent red-team exercises show how traditional controls can fail when safeguards are relaxed in the lab or monitoring does not keep pace.

"Stress-testing frontier AI models for offensive cyber capabilities is exactly what responsible labs should be doing. The problem comes when those tests run with safeguards disabled, weak isolation boundaries, and no effective real-time monitoring. That is when failures spill beyond the lab. AI models optimise for the narrow goal they are given, with no sense of what is out of bounds. They do not need malicious intent to cause harm; they simply need an objective and enough capability to pursue it. This is the heart of agentic risk.

"A human red-teamer knows not to attack a company they were never authorised to touch. An agent optimising a score does not, unless something outside the model stops it. CISOs are under pressure to deploy agentic AI quickly, from their own boards and from every vendor in the market. This is the clearest signal yet that governance has to arrive with the capability, not after it.

"The failure modes here were not exotic: a permission decision, a weak isolation boundary, a monitoring blind spot. Familiar weaknesses, at unfamiliar speed. Authorisation frameworks need to be reimagined to account for autonomous actors. Before deploying an agent, security leaders should be able to answer four questions: What is it authorised to touch? What is explicitly out of scope? Who is alerted when it approaches a boundary? And who owns the outcome if it crosses one? Those answers have to exist before the first deployment, not after the first incident.

"These incidents preview a new category of security risk. The organisations that come through it well will treat autonomous AI as a privileged operator that has to be scoped, watched and owned," said Sprague.

Alongside security concerns, lawyers and compliance specialists point to the EU AI Act's Article 50 transparency requirements, which took effect without the delays granted to other parts of the law under the bloc's Digital Omnibus adjustments.

The rules cover deepfake disclosure, marking AI-generated content in machine-readable formats, and giving users clear information about interactions with AI systems.

Ivana Bartoletti, Global Chief Privacy & AI Governance Officer at Wipro, said organisations must embed transparency into how they design and operate AI products and workflows rather than treat it as an afterthought.

"As the EU AI Act's core transparency obligations take effect this week, organisations should stop treating this as paperwork and start treating it as design. Map the AI systems and content workflows you provide or use, build clear disclosures for deepfakes, machine-readable marking where required, and review processes with real accountability behind them.

"Transparency is not a checkbox; it is a governance choice. It demands technological safeguards, human judgement, education and organisational protocols working together. None of that can be built overnight. The Digital Omnibus deferred the high-risk AI deadlines. It did not touch Article 50.

"While the rest of the Act's timeline moved, this is the part that held, and that should tell businesses something important: postponed deadlines elsewhere do not mean postponed responsibility here. Regulators have extended the timeline on complexity, not on transparency. That distinction is the whole story of this deadline, and missing it is the most expensive mistake a business can make right now.

"Governance by design is no longer optional, and it is no longer just a compliance function. It is what makes innovation scalable, defensible and sustainable. In an era where trust is a competitive differentiator, the organisations that built for this moment will move faster than the ones still explaining why they didn't," said Bartoletti.

Molyneux said boards and technology leaders should reassess their AI security and governance architectures, including how they classify agents, manage access and recover from incidents.

"Following Sunday's ruling, companies with their own AI projects, or those using external AI services, should now assess how far the AI rules apply to them from a governance perspective and how they should rethink their existing models. For IT leaders and CISOs, the task is clear: they need to evolve their security model as quickly as AI adoption advances in their environment.

"A few immutable truths apply. Every AI agent should be treated as a privileged digital identity. Companies should continuously review what an AI agent can access instead of relying on assumptions. Anyone preparing for AI governance needs trusted data and a resilient AI infrastructure. And the most important lesson? Trust in AI must never be taken for granted; it must be continuously verified. Resilience is just as important in enabling rapid recovery, even when the best security controls are bypassed," said Molyneux.