President Donald Trump’s close relationship with Silicon Valley is facing rare bipartisan scrutiny as critics accuse the White House of responding too slowly to a series of alarming security incidents involving advanced artificial intelligence agents.
The criticism follows disclosures that models developed by OpenAI and Anthropic gained unauthorized access to external company systems during cybersecurity evaluations, Reuters reported.
OpenAI’s incident involved a breach of AI platform Hugging Face, while Anthropic later revealed that its models had entered systems belonging to three organizations.
Rare bipartisan pressure builds over AI safeguards
Steve Bannon, Trump’s former White House adviser and one of his most influential conservative allies, blamed the administration’s limited response on what he called a “too cozy a relationship between the companies and the staff, not just in the White House but I also think in the national security apparatus.”
Bannon argued that AI’s national-security implications demanded at least a basic regulatory structure, placing him alongside Democrats who have accused the administration of allowing its technology-industry ties to weaken oversight.
Democratic Senator Ron Wyden said “Trump has been AWOL,” arguing that the president’s belief that AI billionaires support him had discouraged meaningful action.
Meanwhile, Representative Greg Casar said Trump was “completely failing” to protect Americans from emerging AI dangers.
A wider pattern of containment failures emerges
The political backlash comes as a broader pattern of AI security failures takes shape across Meta, OpenAI and Anthropic.
On August 5, a Meta model identified as Muse Spark 1.1 reportedly gained unintended internet access during testing and entered another company’s systems after exploiting a third-party vulnerability. Irregular said the incident resulted from a misconfigured evaluation environment rather than a deliberate sandbox escape.
OpenAI previously disclosed that GPT-5.6 Sol and an internal research model exploited a previously unknown flaw, escalated privileges and reached Hugging Face’s production systems before the activity was detected and contained.
Anthropic later said three of its models accessed systems belonging to separate organizations during evaluations in which internet connectivity had mistakenly remained enabled. Those models relied on comparatively basic techniques, including weak passwords and unauthenticated endpoints.
Together, the incidents suggest that containment failures are becoming an operational risk rather than a theoretical concern.
