Brussels "simplified" the AI Act last week. What actually shipped:
- Sandboxes → 2027 (regulators, not vendors)
- Annex III → still Aug 2026
- Article 50 transparency grace: 6 → 3 months. New deadline 2 Dec 2026, earlier
Three changes, three directions. Only one is relief.
"Saves 2h a week" is an observation, not ROI.
ROI needs: baseline, volume, acceptance rate, end-to-end time, threshold defined before the pilot.
Polarion CR-audit: 6-8h → <2 min. 100% recall, 0 false positives. Measured.
Without those five, you have a story.
April 25, 2026.
A Cursor agent deletes a production database in 9 seconds. Backups too — same volume.
Agent's confession: "I violated every principle I was given."
One of those principles was "NEVER DELETE PROD."
Lesson: system prompts are wishes. Permissions are walls.
Most enterprise agents are just LLM wrappers.
88% of pilots never ship.
Five things they need first:
- Scoped tools
- Enforced rails (not prompts)
- Audit trail per step
- Dry-run mode
- Eval suite in CI
The model isn't the bottleneck in 2026. The agent platform around it is.
86% of enterprise AI pilots never ship. (Gartner, April 2026)
The 14% that do don't have better models. They have:
→ a scoped process with measurable output
→ data access built before the agent
→ ROI defined before the pilot, not after
EU Parliament just pushed AI Act high-risk deadlines to Dec 2027 and Aug 2028. But transparency and GPAI rules still hit August 2026. Only 8 of 27 member states named their AI authority. More time for high-risk. Zero excuse for everything else.
73% of enterprise customers choose Anthropic. In the EU? CLOUD Act hits all US providers. We run agentic workflows on self-hosted GPUs with open-source models. No lock-in. GDPR-compliant by design. Not "best model" — "which one survives the next regulation shift?"
1/4 88% of enterprises report AI agent security incidents.
Only 14% have security approval for their agents.
The gap between those two numbers is where breaches happen.
4/4 Step one is boring but essential: inventory.
You can't secure what you can't see. Map every agent, every data flow, every tool call. Then decide what stays and what goes.
3/4 Prompt injection — manipulating AI through crafted inputs — went from research to production incidents in 12 months.
Most teams treat agents like SaaS tools. They're not. They decide, access data, and act. That needs its own security model.
4/4 This should be automated. Missing test case? Flag it. Outdated after requirement change? Surface it instantly. Not three weeks before the audit.
Traceability should be a byproduct of your workflow. Not the most expensive line item on the assessment invoice.
1/4 ASPICE Level 2 doesn't fail because of bad engineers. It fails because of manual traceability.
Requirements in DOORS. Code in Git. Tests in Excel. Traceability maintained by one person. Manually. Everyone hopes that person doesn't quit.
3/4 Worse: the link exists but the requirement changed. Test case still passes — but tests something that's no longer valid. At 500 requirements, no human catches that manually.
1/4 GPT-5.4 can now operate your computer.
Click buttons. Fill forms. Run legacy ERP systems. Live via API.
Everyone talks benchmarks. Nobody asks: what happens when an AI agent has mouse and keyboard access to your SAP?
4/4 Computer Use will transform enterprise operations. No doubt.
But the companies that win are not the ones who deploy fastest. They're the ones who deploy with guardrails that actually work.
Speed without governance is just a faster way to break things.
3/4 The real question for CISOs: who controls what the agent clicks?
IAM systems were built for humans. Humans don't execute 200 actions per minute or copy customer data into 4 systems at once.
Agents need their own access model. Not shared API keys.