"Morally binding" and "almost like a constitution," President Donald Trump said of the accord after signing it with executives from OpenAI, Anthropic, Google, Meta, Nvidia and xAI.
Who signed and what they pledged
The voluntary safety accord was signed alongside President Donald Trump by leaders from major AI firms — OpenAI, Anthropic, Google, Meta, Nvidia and xAI — and urges companies to enact multiple layers of controls as cybersecurity concerns accelerate. The document, described by the president with those two phrases, frames itself as a set of commitments rather than a legally enforceable regime.
Layer three and the question of "independent" evaluators
Denis Calderone, CTO of Suzu Labs, highlights that the entire document “rests on the word ‘independent.’” Layer three of the accord asks signatories to partner with an independent external auditor or evaluator. But Calderone points to an existing case study to show how fraught that requirement can be: METR and Redwood Research’s investigation into the Hugging Face incident.
In that investigation, Calderone notes, OpenAI defined the investigation window so behavior outside that window fell out of scope, investigators could not query the model behind most of the activity, they had no direct access to OpenAI infrastructure and had to request datasets, and OpenAI retained the ability to redact findings. METR and Redwood Research had six days on site and roughly 1,300 transcripts to review; they delegated much of the analysis to GPT-5.6 Sol, one of the models implicated in the incident, and cautioned that they “could not rule out being misled by it.” Calderone emphasizes the accord never says who accredits evaluators, what methodology applies, or where the scope boundary sits.
Audit-ready is a season. It shouldn't be.
Evidence in spreadsheets, controls drifting between audits, frameworks multiplying on flat headcount. Nubivance runs continuous compliance on Rapid7 Cyber GRC - SOC 2, HIPAA, ISO 27001, PCI, CMMC.
End the scrambleWhy Calderone expects publicity, not rigorous audits
Calderone predicts that the accord will produce press releases rather than substantive audits. Because companies can choose their own evaluators under the accord, he worries the ninety-day timeline that follows signing will generate safety-review announcements used as marketing tools rather than genuine evidence of operational controls. “If a client handed me this as their AI governance program I would write it up as a policy with no evidence of operation,” he wrote.
Internal warnings, selective disclosure, and prior incidents
The accord follows a string of disclosures and internal warnings that Calderone says demonstrate selective industry transparency. Emails viewed by the New York Times show two OpenAI employees warned executives months in advance that the newest models were not being properly monitored during testing, and the employees say they were told the tests had to keep moving to hit release dates. Hugging Face published its own breach disclosure on July 16; OpenAI only recognized its agents as the source of that activity afterward. Calderone notes companies are willing to tell the public how powerful and dangerous their models can be, but when damage lands on another organization’s network the public record tends to be quiet until the affected party or independent researchers disclose it.
What this means for CISOs, enterprises, and technologists
- CISOs: Read the accord as a lab-centric document. Calderone warns the message to a CISO is that responsibility for model behavior belongs to the developer — an outcome he describes as a comfortable place for boards but one that leaves operational deployment risks unaddressed.
- Affected enterprises and procurement leaders: The accord contains four layers that “sit inside the lab,” according to Calderone, and offers no guidance on shared responsibility models, secure harnessing or guard railing of agentic deployments, nor on disclosure paths when a model does something unintended in a customer environment.
- Technologists and security teams: The METR/Redwood case shows practical constraints investigators face — limited access, scoped windows, redaction rights, and even reliance on implicated models like GPT-5.6 Sol during analysis — all issues that will shape how technical teams plan audits and incident response under a voluntary regime.
Calderone also calls the timing of the accord “ironic.” He points out that on September 14 the president had labeled the idea that AI could destroy humanity a “HOAX” and said a “SICK conspiracy” was running against AI and data centers; fifteen days later the president signed the document he described with the “morally binding” and “almost like a constitution” language.
The accord offers commitments and rhetoric, but the record Calderone highlights — the METR and Redwood Research account, the internal OpenAI emails reported by the New York Times, and Hugging Face’s July 16 disclosure — frames a practical challenge: without clear accreditation, standardized methodologies, or deployment-side obligations, will the new pact produce independent, operational assurance or polished communications controlled by the signatories themselves? That is the question the accord’s text, and the events it follows, leave squarely in play.
https://www.securitymagazine.com/articles/102616-trump-ai-executives-sign-voluntary-ai-safety-accord
