AI Agents Just Crossed a Line. Marketing Teams Are Next.

This week, OpenAI said GPT-6 Astra hit a line nobody in the lab wanted to cross first. Critical cybersecurity capability. Translation: with the right tools, the model can find unknown flaws and build exploits without a human walking it through each step.

That is not a product demo. That is a new kind of coworker. And if you run an agency, this is the week the “AI agent” conversation stopped being cute.

What actually happened this week

Three signals landed in the same 72 hours. Treat them as one story.

• Astra is the first OpenAI model marked Critical under their Preparedness Framework. In testing it scored a perfect 100 on ExploitBench and chained two zero-days. Full cyber tools stay gated. The capability does not.

• Researchers say OpenAI-tested agent swarms were implicated in attacks on RubyGems and Hugging Face. Vendor-run agents are no longer a slide. They are in the supply chain.

• On the other side of the aisle, Salesforce shipped named Agentforce agents into production workflows. Enterprise is not waiting for the safety debate to finish.

Same week. Two industries. One conclusion: agents now act. They do not just draft.

The marketing section nobody should skip

While labs argue about pauses, marketing already handed agents the keys.

Forrester says 83 percent of B2C marketing decision makers are implementing agentic AI, and 85 percent say it is delivering real value. Google started converting Search campaigns into AI Max. ChatGPT ads are live enough that dismiss rates dropped in the pilot. IAB buyers are prioritizing agentic systems. GEO is eating the old SEO brief.

McKinsey’s 2026 read is blunt: teams that wire AI into insights, creative, personalization, agentic commerce, and always-on orchestration can see 4 to 7 percent revenue lift and crush execution time. Adobe’s version is harsher. More than 8 in 10 marketing teams missed an opportunity last quarter because they could not move fast enough. Only 7 percent have AI embedded in a way that actually shows up in the numbers.

So here is the agency split I keep seeing.

• LAYER 1 — Speed shops. They let agents write 40 ad variants and call it transformation. Output goes up. Brand gets mushy. Nobody can explain why the winner won.

• LAYER 2 — Operator shops. They give agents a brief, a brand system, a spend cap, and a human veto. Agents draft, test, and assemble. Humans decide what ships and what the brand is allowed to sound like.

We are building Layer 2 at Dzine Prodigy. Not because it is fashionable. Because Layer 1 is how you burn a client’s voice in two sprints.

What to do this week

Do not wait for the next model card. Pick one live campaign. Give an agent the brief, the brand rules, and a hard stop. Make it produce 10 options. You pick two. Ship those. Write down what you refused.

That refusal list is the product now. Models will keep getting faster. Your edge is the taste, the constraints, and the judgment you refuse to automate.

Agents crossed a line this week. Marketing is already on the other side. The only question left is whether your team is supervising them — or hoping they behave.