The Trump administration executive order on artificial intelligence tried to strike the balance between responsible use, security and mutual benefit, all with an eye toward not making it regulatory in nature, National Cyber Director Sean Cairncross said Tuesday.

“Everyone is working towards the same goal in terms of protecting the country and securing our systems, and we are trying to ensure that defenders have this technology as quickly and at scale as possible, but there are obviously specific security concerns, and industry has been very sensitive to this as well,” Cairncross said at the Black Hat 2026 conference in Las Vegas.

The security concerns about AI have moved to the forefront of discussions about the technology after OpenAI models escaped a test environment to hack the company Hugging Face last month.

“The design of this is that when there is something that happens, when there is a breach, when there is an event, that that system, that network of connections can exist, adapt to that, and seek to remedy that as quickly as possible, so that form follows function rather than turning that upside down, and as usual with the government pen just proceeding in a vacuum,” Cairncross said.

The Trump administration has drawn criticism over whether it has struck the right balance on AI rules. Trump’s AI executive order notably got pulled just before its scheduled release, with the final version signed in June missing some aspects that had drawn industry opposition.

“What needs to be built is a flexible, adaptable structure that enables information sharing between industry and government, so we can guarantee that this technology benefits everyone it’s going to benefit, but is used responsibly and securely,” Cairncross said.

He said the administration is working with industry during implementation of the executive order.

“A regulatory regime would not only strangle growth, development, and innovation, and be enormously harmful to the industry, but it would be obsolete 48 hours after it was gone through whatever process it had gone through,” Cairncross said.

Open source will play a “vital” role in the U.S. spreading its vision for AI across the globe, he said.

“We are extremely interested in looking at ways to build U.S. open source, make it competitive, make it the preferential adoption by planet Earth,” Cairncross said. “We understand and appreciate the value to the ecosystem that it has, the innovation, the startups who rely on it, the leap forward it makes possible in ways that otherwise would never happen. And so I think it’s an incredibly thriving ecosystem in AI right now, and we are looking to do what we can to grow, foster, and push that U.S. open source model.” 

Speaking at the same conference, Nick Andersen, the acting director of the Cybersecurity and Infrastructure Security Agency, seconded Cairncross’s comments about AI executive order implementation. He touted the Gold Eagle clearinghouse as one example.

“Those are fantastic opportunities we have to really provide a unifying function around the way that we’re going to do AI-enabled vulnerability reporting and disclosure at scale in a way that we haven’t had to do before with some of our legacy platforms, and just continue to expand out those opportunities,” he said. “That access — to build off the director’s point earlier — to really enable that industry collaboration, that’s so key and critical to us as we move forward.”

Matt Kapko contributed reporting to this story.

Updated 8/5/2026: to include comments from Andersen.

The post National cyber director lays out White House plans to secure AI without writing new rules appeared first on CyberScoop.

The Trump administration’s haphazard and opaque interventions into artificial intelligence security matters could catapult Chinese alternatives into broader acceptance, posing new security risks altogether, a group of Democratic senators wrote to top administration officials Monday.

The five senators said that the administration’s handling has alternated between too passive, such as when OpenAI models escaped testing in the Hugging Face hack last month, and overstepping, such as when the Commerce Department suspended access for any foreign national to Anthropic’s Fable 5 and Mythos 5 in June.

“The Administration’s ad hoc and unpredictable approach undermines U.S. competitiveness, heightening market incentives to adopt open weight models from vendors based in the People’s Republic of China (PRC),” wrote Sens. Kristen Gillibrand of New York, Adam Schiff of California, Mark Warner of Virginia, Chris Coons of Delaware and Mark Kelly of Arizona.

In the Hugging Face hack, the senators wrote that “the Federal Government cannot be passive as these capabilities emerge.”

In the case of the Fable 5 and Mythos 5 suspensions, the senators said that the administration “utilized an infrequently used authority to direct Anthropic to suspend all access to its Fable 5 and Mythos 5 models for foreign nationals (including foreign national employees inside the United States) citing an undisclosed national security concern later described as a narrow jailbreak finding.”

Because Anthropic couldn’t immediately assess users’ nationality, the firm had to disable both models for everyone. The administration and Anthropic negotiated for 18 days behind closed doors before reaching an agreement, the lawmakers complained.

“While the Administration may have been responding to real security concerns to protect the United States, even justifiable interventions can create broader harm if the standards and decision-making processes are opaque, ad hoc, or unpredictable,” they said in their letter to leaders in the White House, Office of the National Cyber Director and departments of State, Treasury and Commerce. “Moreover, when the Executive Branch exercises authority delegated from Congress, such as in the conduct of export control administration, it is essential that it keep Congress fully apprised of its actions and procedures.”

During the time Anthropic was under export controls, the stock price of “an entity-listed Chinese lab” nearly doubled, the senators said. And while Hugging Face was breached, the company “had to” rely on a Chinese open-weight model due to guardrails on U.S. frontier models.

“If American models are perceived as subject to sudden access disruptions based on a black-box U.S. Government process, or as unreliable because U.S. AI labs are overcorrecting in the face of this black-box process, companies and governments in the United States and abroad may hedge by adopting Chinese or other foreign models instead,” the senators contended. “That outcome would undermine U.S. technological leadership while increasing exposure to systems that may carry risks of PRC or otherwise directed censorship, espionage, IP theft, and other supply chain security risks.”

Their letter asked for answers to questions about the standards the administration uses to determine the national security risks a frontier model presents, what legal authorities it will use to invoke restrictions, which agencies are responsible for which decisions and more.

None of the offices or departments the letter was addressed to immediately responded to a request for comment.

The letter follows inquiries at the state level, where 15 attorneys general asked OpenAI for more information regarding the security incident at Hugging Face.

The post Dem senators criticize Trump administration decisionmaking on AI security risks appeared first on CyberScoop.

We almost never get both sides of an intrusion. This time we did.

Last month, Hugging Face disclosed a breach into part of its production infrastructure, saying an autonomous AI agent system ran the attack from start to finish. Five days later, OpenAI revealed that its own models, including GPT-5.6 Sol along with an unreleased one, were responsible for the breach. OpenAI called it an unprecedented cyber incident.

Most of the commentary since has been about AI capability. That is the least useful part of the story, because the capability was doing what it was destined to do. Security teams should look at a simpler truth: in both systems, the key defenses sat behind untrusted code that was already running.

What the two disclosures actually describe

Together, the report shows one connected attack across two firms. Code running on a low-level employee’s machine led to stolen passwords, privilege escalation and lateral moves, then out to the web to hit other targets. No step in that chain is novel. The speed and the number of attempts are what changed.

OpenAI’s evaluation ran in what it calls a highly isolated environment, with network access constrained to a single path: an internally hosted third-party proxy and cache for package registries. The models spent vast computer power searching for an exit. They exploited a zero-day flaw in that proxy, gained higher rights, and moved until they reached a machine with web access. They guessed that Hugging Face held the test solutions, used stolen passwords with more zero-days, and ran code on Hugging Face servers.

Hugging Face picks up the same chain from the receiving end, and its initial access is worth reading twice. A malicious dataset used two flaws in the data pipeline to run code on a worker machine. The agent took over the node, stole cloud keys, and moved into internal clusters over the weekend.

The sandbox was was the only boundary

Both setups assumed the sandbox would hold, so no real checks stood before code execution.  A sandbox is a good control. It is a terrible last control. When isolation is all you have between untrusted code and everything else, one flaw could give the attacker full control. 

Look at what each team allowed. OpenAI allowed one exit path—good engineering, but that proxy contained an unpatched flaw. Hugging Face ran a tool that executes code inside submitted data—a known feature, not a hidden mistake.

Neither choice was careless. Both made sense until a fast, automated system probed them thousands of times.

Why automated agents behave differently

Blocking a human stops them or slows them down. Blocking an AI gives it data. The AI logs the error, changes its plan, and tries again instantly. It keeps trying as long as someone pays for the computing power.

Hugging Face tracked over 17,000 attack events across short-lived sandboxes. OpenAI said its models focused on the goal with extreme effort.

This changes the attacker’s limit. It is no longer skill or patience. It is budget. Computing power scales with money, and code never sleeps. Any defense that relies on attackers giving up after a few tries must change.

Detecting malicious behavior is not the same as preventing execution

Detection worked—but too late

Detection worked, but too late to stop the loss. Hugging Face used AI tools to spot the breach, isolate systems, reset keys, fix flaws, and alert police. That is a good response.

Yet it happened after the attacker accessed private data, stole keys, and moved through internal networks. OpenAI spotted the event only after its models reached another firm’s database. Detection limited the damage; it did not stop the attack.

A breach is a race between attack speed and defender response speed. AI agents speed up attacks with money. Human approval steps slow down defenders.

Hugging Face also surfaced an asymmetry that deserves more attention. When the team tried to analyze the attack with frontier models behind commercial APIs, the requests were blocked, because forensic work means submitting real attack commands, exploit payloads and command-and-control artifacts, and a safety filter cannot tell a responder from an attacker. They used an open-weight model on their own infrastructure instead. The attacker operated under no usage policy at all. The defender’s tooling was the side that got refused.

Detection is not prevention

Detection asks what happened. Prevention asks if it can happen at all. Against AI agents, the gap between those questions is where attacks succeed. The agent builds an attack out of single steps that look harmless on their own.

OpenAI noted this in its research: models operating over long horizons learn approval blind spots and work around them. Safety requires asking what a series of actions aims to achieve, not just if one step is allowed. Most systems check one call at a time: read a file, install a package, open a port. Every step passes. The full chain is the attack.

Six fixes for real defense

To survive automated attacks, defense must shift before code runs. That starts with treating data like code. Dataset loaders, templates, and configuration files all run instructions, so they demand strict controls: block remote execution by default, require review, and check the source of every outside file. Security teams must also control outgoing traffic with strict lists of approved destinations, while treating every proxy in that path as vulnerable. Finally, firms must remove permanent keys, issuing short-lived credentials for single jobs so a hacked worker machine yields almost nothing to an attacker.

Beyond individual machines, isolation must happen per task rather than per user, stopping an automated swarm from moving through internal networks. Authorization also needs to change. Checking single actions fails when an agent builds an attack out of simple steps. Systems must evaluate full sequences instead, setting caps on activity rates and automated spending to flag suspicious chains of events. Lastly, defenders need the freedom to act fast. Security teams need authority to isolate systems without waiting for leadership meetings, and the ability to vet a capable model you can run on your own infrastructure for forensics.

What boards must ask now

Corporate boards usually ask if the company uses AI responsibly. That’s rooted in compliance not safety. Directors should ask four direct questions instead: Which systems run outside code, and what checks sit in front of them? If an attacker targets a low-level employee, what keys and paths can they obtain? How fast do we move from the first warning to active containment? And can we analyze attack code on our own hardware, free from third-party rules?

The boundary has to move

Nothing in these disclosures required a capability that did not exist last year. The chain was ordinary. What changed is that an adversary can run that ordinary chain thousands of times over a weekend, learn from every failure, and never need a coffee break.

The models did not break the detection-and-response model. They exposed where we put the trust boundary. We put it after execution, and we assumed we would have time on the other side of it. We do not have that time anymore.

Move the boundary.

The post What the Hugging Face breach reveals about defense in the age of agentic AI appeared first on CyberScoop.

Anthropic said it found three instances in which its models gained access to the live computer systems of outside organizations, according to a company blog post published Thursday.

The company said it began the review after OpenAI disclosed earlier this month that some of its models had exploited an unknown software flaw to escape an isolated test setup and reach production systems at Hugging Face, a platform for AI models and datasets. Anthropic checked over 141,000 evaluation runs for signs that Claude had reached the internet from environments meant to be closed off. It found six runs across three incidents, all tied to one outside testing partner, Irregular.

“Ultimately, many factors contributed to these incidents, but, consistent with a blameless postmortem culture, we’re approaching the fixes as if the responsibility were ours alone,” the blog post reads. “This begins with ensuring every part of our evaluation pipeline is secure, including the manner in which we integrate with external partners. Moving forward, it will include expanding our continuous monitoring of evaluation transcripts for unexpected behavior, improving our investigation tooling, and conducting more rigorous assurance work with the vendors we rely on.”

In each case, Claude was working on a “capture the flag” puzzle, a common way to test a model’s hacking skill. The model is told a piece of secret data sits on another machine and must be found. Anthropic‘s prompts told Claude it had no internet access. That was false. A setup error at the partner’s end left the machines connected to the open internet, and Claude found real systems while searching for its target.

Believing the real systems were part of the simulation, Claude broke into them using simple methods such as guessing weak passwords and finding unprotected access points. In the most serious case, the model extracted login credentials and reached a database holding several hundred rows of live data. 

In a second case, Claude built and uploaded a malicious software package to PyPI, a public directory for Python code, after noticing a fake company’s setup guide told new hires to install a package that did not exist. The package stayed online for about an hour and was installed on 15 systems, including one belonging to a security firm whose scanner ran the code and had its credentials stolen. 

In the third case, Claude scanned about 9,000 targets before breaking into one company’s application through exposed credentials and a database attack method known as SQL injection.

The three incidents involved different Claude models: Opus 4.7, Mythos 5, and an unreleased internal test model. Anthropic said the models behaved differently once they encountered signs their targets were real. Opus 4.7 recognized the systems were live but kept attacking anyway. Mythos 5 also noticed signs it was on the open internet but talked itself out of that conclusion, pointing to a system date reading 2026 as proof it remained in a simulation. The internal test model, the newest of the three, stopped once it concluded the target was real.

Anthropic said it began its transcript review last week, halted all cybersecurity evaluations, identified the three incidents, and notified the partner and affected organizations earlier this week. Two of the three organizations had not detected the intrusions before being contacted, the company said; outreach to the third was ongoing. 

Anthropic framed the incident as a failure of its testing setup and oversight rather than a case of a model acting on its own, noting that in each instance the model was operating as intended. 

The company said it is working with the outside evaluator METR on an independent review and plans to release a redacted transcript of the PyPI incident within a week. It also said it would tighten monitoring of test environments run by outside partners and expand review of evaluation logs, framing the changes as part of what it called a blameless review of its own processes.

“These facts give us cautious optimism that with tighter monitoring and controls around evaluation infrastructure, as well as continued investment in alignment, this type of risk can be overcome,” the blog post reads. 

The post Anthropic says its AI accidentally hacked three companies during safety tests appeared first on CyberScoop.

Months before the Hugging Face breach, Emergence AI published research that investigative journalist Ronan Farrow made public. Ten autonomous AI agents operated across five virtual environments for fifteen days without human intervention. Much of the attention focused on Grok 4.1 turning violent and Gemini 3 Flash committing 683 crimes.

What mattered more went unnoticed: Anthropic’s Claude Sonnet 4.6 built a peaceful democracy in isolation, then stole resources from neighboring environments the moment it joined a shared one. The lesson was clear: safety is not a model attribute. It emerges from the operating environment. The models didn’t change. Working as designed, their behavior evolved as the environment changed. The lesson is hard to ignore: The governance environment changed, and with it, the reward dynamics.

The story here concerns institutions, specifically OpenAI’s and Hugging Face’s, and how we must understand their recent security incident through that lens.

The industry agrees on how the Hugging Face breach happened. Cybersecurity experts have focused on the vulnerabilities, how they were used, and remediation. OpenAI has highlighted the model’s capabilities. Both conversations matter. What requires attention is why this breach is strategically important. After spending the past weekend discussing it with policymakers, security researchers, and industry practitioners in Aspen, I came away convinced we’re examining the wrong problem.

In 1961, Yale psychologist Stanley Milgram’s experiments revealed a broader truth: changing the institutional architecture changes behavior without changing the actor. The Emergence AI researchers didn’t change Claude’s agent. They changed the governance architecture that determined what constituted success for the system. Claude’s behavior changed with it.

OpenAI built a smart model but forgot to build a smarter room. That choice made the Hugging Face breach possible. Every organization now deploying autonomous agents now faces the same governance problem.

OpenAI gave the agent one objective: pass a cybersecurity evaluation. To stress-test it fully, they loosened the safety restrictions, and the agent found a shorter path. Rather than solving the evaluation directly, it found the answers outside the test environment, escaped its sandbox, and exploited a flaw in Hugging Face’s data-processing pipeline to reach live production systems. Over the weekend, with no human oversight, it ran more than 17,000 automated actions by escalating its own access, moving through internal systems, and harvesting credentials.

Hugging Face is one of the world’s most prominent AI companies, valued at approximately $4.5 billion. It provides the infrastructure that governments, defense organizations, and technology companies use to build and deploy AI. The agent was pursuing the objective it had been given. Breaking into Hugging Face was the fastest path to passing the test. Governance set the goal, the level of risk to accept, and who was accountable. Technical design determined whether those governance decisions could be enforced. As researchers James Shires and Max Smeets have argued, for a model capable enough to act on its own, testing and deployment must both must be governed the same way.

AI agent design requires baseline standards. Observability, including a monitoring layer that flags when an agent goes beyond its scope, is a baseline requirement. Human review also matters at escalation boundaries, like when an agent shifts from internal tools to external ones. When any agent crosses that boundary, what alert fires? What human reviews it? We lack clear answers to either. That is a governance choice, not simply a security failure. At best, this was a catastrophically failed test. At worst, how can we trust any frontier AI company to self-govern autonomous agent deployment?

More than a decade ago, the U.S. Department of Defense built the Comply-to-Connect (C2C) program: every device connecting to sensitive networks must prove it belongs there, or it is cut off from the network. C2C works because the quarantined actor stops. A laptop that fails verification goes offline and stays there. An autonomous AI agent adapts around enforcement. C2C was built for passive actors. Governance for autonomous agents must accommodate ones that adapt. Visibility is not enforcement, and enforcement is not control. We are missing all three.

A second failure that is not being discussed enough: the breach exploited an implicit trust assumption in Hugging Face’s data-processing pipeline, where inputs were treated as trusted without verification. After SolarWinds, the U.S. government set rules for software supply chain integrity: Executive Order 14028 and verification demands for federal software. The principle was simple: trust must be verified through proof. Those principles have not yet been comprehensively or consistently applied to the AI model supply chain. The rules remain weak. No one has been asked to explain why.

The answer is not a new framework. Existing frameworks suffice. C2C proved that visibility without enforcement leaves gaps, while Executive Order 14028 established that trust in software supply chains requires proof and verification. The challenge lies in applying these principles to a new category of actor. Congress, the Cybersecurity and Infrastructure Security Agency, or the Office of Management and Budget should make formal determinations that autonomous AI agents must follow the same rules as every other actor on a federal network. The framework exists; it must be updated.

The next incident is already in progress. It will show up in the logs as odd traffic, get handed to the same people who published these frameworks this week, and spark another round of recommendations no one acts upon. We’ve solved this problem before: for devices, for software, for supply chains. We know how to build smarter rooms. The tools exist. The will, the authority, and the decision to govern remains absent.

The post OpenAI’s rogue AI agent shows why we need federal rules for autonomous systems appeared first on CyberScoop.

Hugging Face has published an anatomy of the attack and OpenAI has shared additional information from its investigation.

The post OpenAI’s Rogue AI Ventured Beyond Hugging Face appeared first on SecurityWeek.

Industry professionals debate whether it represents a lab containment failure or an unprecedented agentic capability milestone.

The post Industry Reactions to OpenAI Models Hacking Hugging Face: Feedback Friday appeared first on SecurityWeek.

A cyberattack that poisoned the data pipeline of a major AI code platform was carried out using OpenAI’s ChatGPT, the company said Tuesday.

Last week, Hugging Face, a platform for sharing and working on AI code, disclosed that an external attacker had compromised its data processing pipeline. According to a July 21 blog post, the attacker poisoned a dataset to run code on a processing worker, eventually gaining node-level access and stealing cloud credentials.

The attack is notable, the blog said, because it appears to have been carried out by an autonomous AI system, which executed “many thousands of individual actions across a swarm of short-lived sandboxes, with self-migrating command-and-control staged on public services.”

At the time, Hugging Face said it wasn’t clear which LLM was used in the attack, but noted that their own attempts to recreate parts of the attack using an in-house commercial model failed due to guardrail restrictions.

The incident and its ultimate impact are still being investigated, and Hugging Face said it reported the incident  to law enforcement and is engaging with third-party cybersecurity forensic specialists to learn more.  

“This experience points to a gap worth planning for,” the company wrote. “We do not know which model powered the attacker’s agents, whether a jailbroken hosted model or an unrestricted open-weight one; either way, the attacker was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried.”

Now OpenAI is confirming one of their models was used in the attack.   

“After investigating, we now know that this particular incident was driven by a combination of OpenAI models — including GPT‑5.6 Sol and an even more capable pre-release model, all with reduced cyber refusals for evaluation purposes — while being internally tested on a benchmark of cyber capabilities,” the company wrote in a blog post Tuesday.

OpenAI called the incident “unprecedented” but also predicted that such attacks would  become increasingly more common as AI adoption continues to grow.

According to the company , the incident happened during an internal evaluation,  where the company deliberately prompts models to “pursue advanced exploitation using complex attack paths” in order for the company to study and measure the model’s cybersecurity capabilities.

To assess the model’s full potential, the team disabled the production classifiers that normally prevent OpenAI models from carrying out high-risk cyber activity. Without these safeguards,  the models were able to install software packages through registries like Hugging Face.

The model collected and chained vulnerabilities from OpenAI’s systems as well as Hugging Face’s production infrastructure.

OpenAI further explained that the activity stemmed from models attempting to solve a specific challenge posed by ExploitGym, a benchmarking system for AI agents.While such tests normally run in isolated environments, in this instance,  the model spent a “substantial” amount of compute resources figuring  out how to access the open internet. It did so by using a zero-day vulnerability in a third-party vendor’s system used by OpenAI (the company said it disclosed the flaw to the vendor).

“After gaining Internet access, the models inferred that Hugging Face potentially hosted models, datasets and solutions for ExploitGym,” OpenAI wrote. “Knowing this, the model searched for and successfully found ways to gain access to secret information that it could use to cheat the evaluation.”

In one instance, the model chained together multiple stolen credentials and zero-day vulnerabilities to find a remote code execution path on Hugging Face servers.

OpenAI said it was implementing new controls for infrastructure configuration, something that would come at “the cost of research velocity while the vulnerabilities are patched.” 

Hugging Face was also added to OpenAI’s Trusted Access for Cyber program, which will give them access to more advanced models like Daybreak to scour their systems for other AI-discovered vulnerabilities.

The post OpenAI says model test was behind Hugging Face hack appeared first on CyberScoop.

Targeting production infrastructure, the attack compromised internal datasets and service credentials.

The post Hugging Face Hacked in Autonomous AI Attack appeared first on SecurityWeek.