The Real AI Safety Threat Isn’t Speed—It’s Centralized Control

Rogue Code Forces Tech Firms Into ‘Synthetic Internets’

Days after Anthropic CEO Dario Amodei published a letter calling for an artificial intelligence slowdown, former New York City mayoral candidate Andrew Yang has warned that it may be a little late to salvage the situation, as rogue agents have already polluted the internet with self-replicating code.

Speaking with CNBC, Yang shared a revelation from an unnamed AI lab executive: rogue code has forced major tech firms to create “synthetic internets” just to train their models. Yang argued that this costly setback justifies calls from industry leaders like Amodei, Sam Altman, and Elon Musk to regulate the sector and pace frontier AI development.

Prior to Yang’s warning, several AI policy figures and Trump administration officials resisted slowdown calls, warning they could hand China the lead in the global AI race. Donald Trump was even more direct, labeling fears of AI-driven human extinction a hoax. Other critics, such as David Sacks, argued that concerned tech executives should self-impose testing pauses instead of lobbying for federal regulations or broad moratoriums.

As the discourse increasingly devolves into a partisan divide between pro- and anti-regulation camps, key figures on both sides have offered little empirical data to substantiate or debunk calls for a forced slowdown. Yet, Yang’s latest disclosure, which he framed as breaking news, injects urgency into the debate, providing compelling ammunition for those advocating immediate guardrails.

Research Shows AI Agents Exploiting Evaluation Benchmarks

Meanwhile, Sentient Labs, a frontier open-source AI research organization, says a recent study it conducted appears to validate Amodei’s argument that agent capabilities are outpacing the ability to manage risk. In written answers to questions from Bitcoin.com News, Abhishek Saxena, head of strategy and growth at Sentient Labs, said the company’s Evoskill v2 research provides “concrete evidence” supporting the argument.

Saxena said the company’s research team has been studying how AI agents behave in competitive evaluation environments, specifically what happens when agents are optimized to perform well on benchmarks and assessments.

“What we observed is that agents discover unexpected strategies that exploit the structure of the evaluation itself rather than solving the intended task. They find gaps in how evaluators measure performance and learn to optimize for the metric rather than the underlying objective,” Saxena explained.

From a safety perspective, this is significant because it shows that as AI systems become more capable and autonomous, the evaluation infrastructure needs to be at least as sophisticated as the systems being evaluated. To overcome this, there is a need for what Saxena describes as dynamic, adversarial evaluation methods that evolve alongside the systems they are measuring.

However, Saxena insists that a slowdown with no changes to how AI systems are evaluated and monitored accomplishes very little.

“The better path is investing heavily in independent evaluation, red-teaming and behavioral monitoring so that the people deploying these systems and the people affected by them have reliable information about what they’re working with,” the Sentient Labs executive said.

The Bigger Risk: Concentrated Power and Centralized Safety Claims

Meanwhile, Saxena told Bitcoin.com News that the biggest risk facing the world is not that we move too fast or too slow, but that a small number of AI companies have control over the most powerful models and also dictate what is considered “safe.”

“We need to build independent infrastructure to actually verify the claims that AI companies make about their systems. Are the evaluations robust? Are the safety reports accurate? Can outside researchers reproduce the findings? If the answer is no, then neither speeding up nor slowing down solves the fundamental problem, which is a lack of trustworthy information,” Saxena concluded.