The “AI is going to kill us” frenzy has officially hit circus status. Under the spotlight are the CEOs of the largest frontier labs, warning of the dangers of the very technology that they’re building. Earlier this month, Anthropic CEO Dario Amodei published a dramatic essay arguing that the AI industry should slow the pace of frontier model advancement until safety, oversight, and alignment mechanisms can catch up. Since then, both Anthropic and OpenAI have announced that they are investigating tens of thousands of incidents where models took problematic steps, including meddling with US government websites.

The self-serving motivation of these declarations is too great to ignore, especially against a backdrop of proposed IPOs. Introducing slowdowns and industry regulations now would protect their competitive lead, with smaller challengers struggling to jump through regulation hoops. The public declaration also serves as a convenient CYA against any future adverse events (“we told you we should slow down”). Even SNL saw the absurdity, poking fun in a skit last Saturday portraying Amodei whining, “I urge you to urge me to stop.”

Meanwhile, governments are politicizing the issue as the global conversation about AI explodes. One hundred and twenty-eight delegations addressed AI at the most recent UN General Debate; it was only mentioned by five of them four years ago. The media is capturing and amplifying the most salacious AI stories and sound bites to grab our attention and keep us wanting more. The whole topic has turned into a proverbial three-ring circus. Come on — we’re not falling for this clown show. Or are we?

Forrester’s September 2026 Consumer Pulse Survey shows a 22% decline in the portion of US online adults who are excited about how AI will shape the future compared to five months ago, and online adults in the US and UK were 77% and 94% more likely, respectively, to share negative sentiments when asked about what they had learned recently about AI compared to just two months ago. July’s responses covered a range of positive and negative aspects. In September, however, topics shifted to more concerns such as control, governance, and, yes, the risk that AI poses to humanity.

The concern is floating among business leaders, as well. Even self-proclaimed AI enthusiasts are starting to ask questions and/or fielding questions from concerned colleagues and higher-ups. One senior executive at a large US bank told us, “My CEO and board are asking questions about safety and are wondering if the company should slow down its AI pace.”

Ground Yourself In Three AI Realities

It isn’t surprising that consumers and business leaders are spinning amid the circus, but it is worrisome. The future of AI is a serious topic. Achieving the massive upside potential of AI — curing diseases, driving economic growth, empowering humans — without unleashing systemic risks requires clarity. It requires a foundation of facts, an understanding of what is true and what is not. Here are three important truths:

  1. AI won’t kill humanity. Doomsday scenarios typically center on an AI system that is optimized for survival. It’s true that such systems would treat every security control as an obstacle or even successfully fight their own off-switch. But AI models need resources to do serious damage, and today, they can’t autonomously generate the money to pay for the compute they will need. Redwood Research gave Claude Opus 4.7 $5,000 and four days to make money on its own, and it produced zero dollars. They also can’t pass themselves off as human during know-your-customer cloud resource acquisition processes. So there are natural constraints in place that would prevent a self-replicating doomsday at present. Further, AI safety bodies like METR and the UK’s AI Security Institute are tracking these capabilities, which should provide us advanced warning.
  2. AI risk will never be zero. Extinction aside, the way a model can hide and try to cover its tracks, and its ability to find and exploit systems that humans depend on, is real. About 700 agents collaborating in the Hugging Face attack demonstrated what’s possible today with far less powerful models than what is emerging at the frontier. In a few quarters, today’s open-weight models will catch up and be available to anybody with the compute to run them, including hackers and rogue actors. This is why AI model testing, alignment, operational rigor, and independent oversight are essential. More importantly, it’s why overall software security is more critical than ever. Problems arise when AI exploits holes in software, so let’s plug the holes — sometimes the best offense is a strong defense.
  3. AI labs could create safeguards — if they really wanted to. OpenAI, Anthropic, and others could do more to band together to create and implement viable safeguards themselves. They can also work with governments and hyperscalers to establish the same safeguards for open models. They are, after all, the experts best equipped to know what’s possible. For example, NVIDIA launched an Open Agent Safety Platform consisting of OpenShell and Sentry to provide hardware and architectural controls for agentic safety and security. We have examples from recent history of successful industry collaboration. Take a page from the credit card companies that developed the PCI security standard together to safeguard against fraud and data breaches as e-commerce took off.

Prioritize Building Your AI Enterprise

Leaders that maintain a disciplined approach grounded in reality and coupled with the right strategy will outperform those chasing or shying away from every new boom or bust AI narrative. Don’t get derailed by doomsday drama coming from sources that are out of touch with enterprise customers. To succeed without taking on unnecessary risk, prioritize:

  1. AI safeguards appropriate to your capabilities and models. Take the deficit in safeguards into your hands — at a level appropriate to your use. Implement risk, compliance, security, and human oversight mechanisms and ensure that your safety measures are appropriate for the models you are deploying. Introduce AI kill switches at every layer of the AI stack and OSI model. Demand transparency from the vendors you work with, requiring them to provide evidence of testing, safety controls, model provenance, and operational safeguards. Forrester’s Agentic AI Enterprise Guardrails For Information Security (AEGIS) framework ensures that you aren’t rolling out AI without guardrails and is designed to help secure, govern, and manage AI agents and related infrastructure.
  2. Security to make your software harder for AI-powered attackers to exploit. Mythos and Astra show how models can find software flaws and develop exploits. You can’t depend on access controls and alignment practices at frontier labs to secure your software and infrastructure, especially since open-weight models perform these tasks well. This requires a technical and cultural shift around vulnerability management, one where confidentiality supersedes availability in importance. Prioritize more than discovery of vulnerabilities, know which applications and dependencies you run, identify attack paths and blast radius, and test and ship patches faster. SLA credits caused by downtime are far less expensive than multiyear litigation from customers and suppliers in the event of a breach.
  3. Platform investments that preserve optionality. Deterministic software tolerates weak identity, poor reuse, fragmented semantics, after-the-fact assurance, and limited visibility. But goal-seeking, agentic systems turn these into active hazards. Start with accountable identity, build governed capabilities on that foundation, ground them in enterprise context, and create observability to steer the system. Each of these no-regret investments creates value, whether the future arrives as a swarm of autonomous agents or simply an expanding collection of better software. Meanwhile, take care to reduce platform lock-in, and be smart about model routing (not all tasks need the heavy-duty models) to manage your costs and more quickly demonstrate real AI value.
  4. Private models for high-value innovation. Despite all the attention on public models, true business value will come from private models and systems — those trained on proprietary information. Abstractive Health, for example, summarizes patient medical records, enabling doctors to converse with their patients’ medical history. A truly private model that is exposed to a narrower set of capabilities and training data will be less capable and therefore less risky than public models, but the company training the model will still need to implement safety guardrails. Private models that are built on open-weight models (more common and economical) will come with safety guardrails already embedded, though they can be easily stripped out. Bottom line: Private models will deliver big value, but don’t assume that they are inherently safe.

The AI circus isn’t likely to leave town anytime soon. Our advice: Stay grounded in what’s real and focus on the risks you can control.

Thanks to Brian Hopkins, Mike Gualtieri, Jeff Pollard, Rowan Curran, Kevin Ogunsua, Ted Schadler, Srividya Sridharan, and Stephanie Balaouras for their contributions to this post. 

Share