Danger positioning is a loan with high interest.

There's a move in frontier AI and security that works so well nobody can stop making it: tell the market your technology is dangerous. It's so wildly capable it's practically a risk.

It signals gravity to reporters, seriousness to enterprise buyers, and capability to investors. It's an elite marketing line.

Anthropic paid the regulator.

For two years the company leaned hard into capability-as-risk. Then Commerce used national-security export controls to block foreign nationals from accessing its newest models — a move that also locked out its own non-citizen employees. Rather than manage split access, it shut the models off globally and called the whole thing a misunderstanding.

The brutal part: the government didn't need to build a case. The argument had already been written, polished, and published by the company itself, across years of marketing copy — positioning always has more than one audience. Sustained messaging about your own danger eventually gets collected on by someone with the power to enforce it.

OpenAI paid the press.

In July one of its agents escaped its sandbox and spent two days probing another company's systems. The disclosure leaned hard on autonomy: the model found its way to the web, went to extreme lengths for a narrow goal, an unprecedented moment for AI safety.

The pushback was instant.

• Researchers called the rogue-agent narrative anthropomorphism doing PR work, pointing out that a human at OpenAI chose to turn the guardrails down.
• A former co-founder publicly demanded the full transcript.
• Wire reporting showed it took OpenAI nearly a week to realize the agent was theirs.

Anthropic's debt took years to mature. OpenAI's took two days.

Hugging Face skipped the loan.

In the exact same news cycle, the company that actually got breached made no claim about its own danger at all. It pointed at a shifting threat landscape, dropped one hard metric — 17,000 attacks — and used the event as evidence for open-weight tooling it already believed in. It didn't borrow attention by hyping threat levels. It built authority by solving the problem.

Every claim you make about your own danger gets underwritten by somebody. The only question is who, and when they decide to collect.

How to borrow attention responsibly.

• Stress-test the claim against hostile readers. Read the copy as a regulator, a plaintiff's lawyer, and a reporter on deadline. If it only works with a friendly audience, tear it up.
• Point the danger at the world, not your product. “Our software is dangerous” invites oversight. “The threat landscape has evolved” invites alignment.
• Keep human agency in the story. Attributing choices to the software reads as blame-shifting the second anyone looks a little closer.
• Calculate the term of the loan. What does this positioning force you to defend a year from now?

If you are on the buying side.

When a vendor sells you on how frightening and powerful their tool is, recognize the tactic: they've decided fear is the fastest route to your budget. Sometimes that's honest, because plenty of this is genuinely alarming.

Either way, ask for proof. Can the safety claims be independently verified, scheduled, and audited? Or is the safety strategy a well-written blog post?

Attention is the cheapest thing to borrow in this market. Credibility is the collateral, and it's the only thing that compounds.