The EU AI Act compliance clock is ticking.
Staggered enforcement timelines are underway, with potentially sweeping liabilities entering the picture. A high-risk compliance deadline of December 2027 is approaching faster than most teams expect, exposing organizations operating in the EU to a new kind of regulatory pressure.
This is not just compliance paperwork—organizations are facing real accountability.
But there's a bigger problem hiding beneath the regulatory noise: the technology most organizations are rushing to deploy—agentic AI—is fundamentally different from anything the governance playbooks were written for. And most security, compliance, and IT teams aren't ready for it. What will it take to prepare your organization for the coming changes?
What the EU AI Act Actually Requires
The EU AI Act is the world's first comprehensive, horizontal legal framework for AI. It passed the European Parliament in March 2024, was fully endorsed in May 2024, and began coming into force in August 2024. Obligations have been coming into effect on a staggered basis ever since.
The Act takes a risk-based approach across four tiers:
Unacceptable Risk includes social scoring, manipulation of democratic processes.These are banned outright.
High Risk includes things like biometrics, critical infrastructure, employment, healthcare, law enforcement.
Limited Risk includes transparency obligations.
Minimal Risk which requires no mandatory compliance for most uses.
One critical point for non-EU companies: the Act is extraterritorial. If your AI systems touch the EU (processing EU citizen data, operating in EU markets, etc.) it may apply to you regardless of where you're headquartered.
For providers of high-risk systems, obligations include conformity assessments, risk management systems, documentation standards, and EU registration. For deployers, that means following provider instructions, ensuring human oversight, monitoring operations, and retaining logs for a minimum of six months.
While December 2027 gives some breathing room on the heaviest obligations, provisions requiring employee AI literacy are already in effect.
So the runway is shorter than it looks.
The Agentic AI Shift Changes Everything
The urgency isn't just regulatory. It's technical.
According to McKinsey, 62% of enterprises are already experimenting with or scaling AI agents. The World Economic Forum reports that 87% of executives rank AI as the fastest-growing cyber risk. We've moved, very quickly, from chatbot-based systems that answer questions to agentic systems in which large language models act as a "brain," with tools and decision logic as "limbs," to interact with the world autonomously.
These systems are non-deterministic. They can hallucinate. And they can go significantly out of bounds.
Consider what "out of bounds" looks like in practice: an agent connected to a mailbox, tasked with optimizing email organization, might accidentally delete hundreds of emails. Not because it was attacked, but because it interpreted its objective too liberally.
The Governance Gap Is Not Theoretical
Real incidents are already on the record. An AI coding assistant executed commands that deleted two production websites and their backups, wiping out 2.5 years of data. A separate AI coding agent from Replit deleted a live database during a code freeze. At Meta, the Director of AI Safety had her inbox wiped by an agent she was testing.
These aren't hallucinations. They are executed actions with real-world consequences.
Research from Rubrik Zero Labs documents an even more troubling pattern: identity spoofing works. An attacker who simply changed their display name to match a system owner gained full workspace takeover with zero cryptographic verification. Refusals are trivially reframed: an agent that refused to "share" emails containing SSNs immediately complied when asked to "forward" them instead. And emotional persistence breaks agents: after 12+ refusals, sustained guilt-tripping eventually caused the agent to comply. It appears social engineering has been adapted for AI.
If your security posture assumes your agents will behave as instructed, these findings should give you pause.
Why Static Rules and Human-in-the-Loop Don't Scale
The instinctive response to agentic risk is to add human checkpoints or write more guardrails. Both interventions break down almost immediately.
Agents operate at machine speed. They chain tool calls in what researchers call Sequential Tool Attack Chaining (STAC), where each individual action looks legitimate but the combined sequence achieves something harmful. By the time a human reviews step one, steps two through fifteen have already executed.
Static rules face a different failure mode: they become stale. LLMs update, policies evolve, regulations change: but a static rule that says "don't discuss financial advice" won't catch an agent that offers to "add something to a portfolio." The request may use different words, but the intent (and impact) is the same. Traditional keyword-based rules can't reason about intent.
Assume Overreach
Cybersecurity has long operated on an "Assume Breach" principle—design your defenses assuming a malicious actor has already gotten in.
Agentic AI demands an analogous shift: Assume Overreach.
Assume that at some point your AI agents will act beyond their intended boundaries, not because they're malicious, but because they're probabilistic. An agent that's 99% reliable will still fail thousands of times at enterprise scale.
The response to Assume Overreach isn't fear. It's external enforcement.
Operational Control: Using AI to Control AI
Rubrik's approach centers on a principle that sounds counterintuitive at first: use AI to control AI.
Rather than hardcoding rules that agents must self-police, the model uses the Semantic AI Governance Engine (SAGE), a small language model that sits externally to the agent and evaluates behavior in real time. SAGE reasons about intent rather than matching keywords, so it catches policy violations regardless of how they're phrased. A natural language policy like "the agent is not allowed to provide financial advice" will flag "add this stock to your portfolio" just as reliably as "buy this stock."
Rubrik Agent Cloud operationalizes this across three pillars:
Visibility: Discover all agents, tools, and data interactions across your environment
Governance: Build and apply policies at runtime in plain English—no static rules, no code required
Remediation: Prevent dangerous activity before it executes or use Agent Rewind to recover data and state when an agent makes a mistake
This maps directly to what the EU AI Act requires from deployers: human oversight, operational monitoring, intervention capability, and log retention. It's not just a security product, it's a compliance control plane for the agentic era.
What Security, IT, and Compliance Teams Should Do Now
The EU AI Act's obligations and the practical risks of agentic AI aren't separate problems. They're the same problem viewed from different angles.
Start by mapping your agentic footprint: audit which agents are operating in your environment, what tools they have access to, and what data they touch. Many organizations don't know.
Then shift from static rules to adaptive enforcement: runtime control that reasons about intent is the only mechanism that scales for non-deterministic systems. Finally, build for recovery, not just prevention—point-in-time rollback capabilities need to be in place before the incident, not after.
The December 2027 deadline feels distant. The agents causing damage are deployed today.
Watch the full webinar Before the Deadline: Governing Agentic AI in the EU AI Act Era featuring Neal Burstyn and Filip Verloy.
Download the Rubrik Zero Labs report on agentic AI risk.
Learn more about Rubrik Agent Cloud and SAGE.