Concerns over new security gaps in artificial intelligence are pushing policy advisers aligned with former President Donald Trump to revisit a light-touch stance on the technology. The shift follows fresh findings from Anthropic’s system known as Mythos, which uncovered hidden weaknesses that could be exploited by attackers. The discussions signal a possible turn in how a future Republican administration might set guardrails for powerful models.
“As AI systems like Anthropic’s Mythos expose hidden security flaws, Trump officials are starting to rethink their hands-off approach to the technology.”
People close to the talks describe early-stage debates on testing, disclosure, and procurement standards. The timing reflects rising concern across parties that advanced models can be misused or manipulated, while still powering large parts of the economy.
From Light-Touch to Safety Checks
During Trump’s first term, the White House promoted U.S. leadership in AI through a growth-first lens. A 2019 executive order sought more federal research funding and workforce training, while signaling skepticism of new rules. That approach mirrored a broader push to limit regulation in fast-growing tech sectors.
Since then, models have grown in scale and reach. Companies deploy chatbots in customer support, code generation, and search. Governments use AI in fraud detection and cybersecurity. As use spreads, so do risks, including prompt injection, model jailbreaks, data leaks, and the spread of convincing misinformation.
Under the following administration, federal agencies advanced voluntary safety commitments and a risk management framework through the National Institute of Standards and Technology. Industry red-team exercises became common. Now, the new security findings cited above are adding urgency to Republican policy circles long wary of heavy-handed rules.
What the Flaws Mean
Mythos’ results point to ways models can be steered into revealing sensitive information or producing harmful outputs. Security testers often try to bypass safeguards by chaining prompts, translating instructions, or leveraging external tools. Even small failure rates can have large effects when models sit inside business workflows.
Experts warn that attackers can automate these exploits. That raises stakes for companies that embed AI in finance, health, and critical infrastructure. It also exposes gaps in vendor security claims and in buyers’ due diligence.
- Prompt injection can override safety rules.
- Data leakage may reveal training data or private inputs.
- Tool misuse can trigger real-world actions with minimal oversight.
Policy Options Under Discussion
People briefed on the conversations describe a set of ideas that stop short of broad regulation but add targeted guardrails:
- Independent testing before federal adoption of large models.
- Incident and vulnerability disclosure requirements for vendors serving government.
- Procurement standards tied to NIST guidelines and regular red-teaming.
- Accountability measures for deploying AI in high-risk settings, such as critical infrastructure and elections.
Republican-aligned advisers are also weighing export controls on advanced model weights if national security is at risk. Supporters say these steps protect innovation by reducing the chance of high-profile failures that trigger broader backlash. Critics warn that new rules, even if narrow, could slow product cycles and entrench large firms that can afford compliance.
Industry and Expert Views
Safety researchers have long called for testing standards similar to those in aviation or pharmaceuticals, scaled to AI’s unique risks. They argue that independent audits and clear benchmarks can catch flaws early. Companies, for their part, point to rapid security patches and voluntary commitments as proof they can manage risks without strict mandates.
Civil liberties groups urge caution. They worry that poorly written rules could expand surveillance or restrict open research. Small developers caution that excessive compliance costs may push them out of the market, reducing competition and slowing open innovation.
What to Watch Next
The near-term focus is likely to be federal use of AI. Procurement rules can move faster than new laws and shape market behavior. Agencies could require model evaluations under standard threat scenarios, verified by third parties. Lawmakers may also hold hearings to measure the need for incident reporting and safe release practices for powerful models.
State officials are moving as well, considering bills on AI disclosures in elections and consumer protection. Any federal shift from a hands-off approach to targeted safety checks would set a national baseline, even if narrow.
The rethinking among Trump-aligned officials suggests a new consensus may be forming: keep innovation strong, but set minimum safety bars where the stakes are highest. The coming months will show whether those ideas harden into policy. Watch for procurement updates, testing standards, and coordinated guidance from security agencies. If model flaws keep surfacing, the push for clearer rules is likely to grow.
