ai risk today logo svg bl font
Menu
SUBSCRIBE
SUBSCRIBE
ai risk today logo svg bl font
Menu

cybersecurity

Abstract AI incident reporting image with glowing nodes and warning patterns for the AI Incident Reporting Act.

AI Incident Reporting Act Would Require Frontier AI Developers To Report Critical Incidents

, , , , , , , , ,

The proposal from Rep. Nathaniel Moran would give covered model developers seven days to report serious AI incidents to the Commerce Department.

Rep. Nathaniel Moran (R-Texas) introduced the AI Incident Reporting Act, a bill that would require developers of the most advanced AI models to report dangerous capabilities, security breaches, and safety incidents to the Department of Commerce.

The bill would apply to “covered model developers,” a category Commerce would define through capability thresholds for AI models that could pose significant risks to national security or public safety. Developers covered would have seven days to report critical incidents.

The legislation would require reporting when an AI model exhibits signs of rogue behavior, including evading human oversight or undermining human control.

It would also require reporting for several other categories of high-risk activity, including:

  • Unauthorized access to, theft of, or attempted theft of model weights (the numerical parameters learned during training that determine how an AI model processes inputs and generates outputs)
  • AI capabilities that could enable cyber attacks against critical infrastructure or widely used systems
  • Evidence that a model can accelerate or automate the development of more advanced AI systems
  • Capabilities that could enable the development or use of chemical, biological, radiological, nuclear, or explosive weapons
  • Serious near misses where harm was avoided because of outside circumstances, not the developer’s safeguards

For incidents that present an imminent or ongoing risk of serious harm, Commerce would have to notify congressional leaders and relevant committee chairs within 48 hours.

Commerce would have to issue reporting guidelines within 180 days after enactment. The bill also directs the department to create a secure reporting mechanism for AI companies.

“AI is a powerful engine of innovation, and I want to see it flourish, but not without accountability and not without human oversight,” Moran said in the press release. “The rule of law should apply to this new frontier.”

The bill also includes protections for sensitive information submitted to the government. Reports would be exempt from public disclosure, would not waive trade secret or attorney-client protections, and generally could not be used by federal, state, or local governments to regulate or bring enforcement actions against the developer.

Commerce could act based on reports to address risks to national security or public safety. The bill would also allow Commerce to share information with other agencies, including intelligence and law enforcement agencies, as appropriate.

The bill does not yet appear to have cleared committee.

Anthropic CEO Dario Amodei standing before the White House, representing the company's cybersecurity collaboration with the U.S. government through Project Glasswing

U.S. Government Used Anthropic’s Mythos To Find Classified System Vulnerabilities Within Hours

, , , , , , , , , , , ,

The government participates in Project Glasswing, which gives vetted partners access to Mythos, the Anthropic model that has found more than 10,000 software vulnerabilities since April.

 

The U.S. government used Anthropic’s Mythos AI models to identify vulnerabilities in classified systems within hours, according to the Associated Press. It is the most sensitive disclosed application yet of Project Glasswing, Anthropic’s restricted program that gives vetted organizations access to its most powerful AI models to find software vulnerabilities before attackers do.

It is unclear whether the government used Mythos Preview, the original version launched in April, or Mythos 5, the latest version. The U.S. government suspended access to Mythos 5 on June 12 after citing a jailbreak identified in Fable 5, the consumer version of the same model, and halted access to both on security grounds.

Anthropic launched Project Glasswing, concurrently with Mythos, after determining that Mythos was too effective at finding vulnerabilities to release publicly. The company restricted access to a short list of vetted technology and cybersecurity companies, along with select government organizations. Partners found more than 10,000 high- and critical-severity software vulnerabilities in the program’s first weeks.

The government and Anthropic continue to work together despite the ongoing conflict between the parties. In March, the Department of Defense designated Anthropic a supply chain risk after Anthropic requested assurances its models would not be used for mass domestic surveillance or fully autonomous weapons. Anthropic challenged the designation in federal court and secured a temporary injunction. In June, the government separately ordered Anthropic to suspend access to Fable 5 and Mythos 5 due to concerns about access by foreign nationals. Anthropic complied while publicly disputing the evidence behind the order. The lawsuit remains active.

The White House held direct talks with Anthropic CEO Dario Amodei earlier this year about cybersecurity and shared AI priorities. Anthropic said it plans to expand Glasswing’s membership.

Blurred photo of a computer monitor showing dense multicolored code, from the UK government's AI vulnerability scanning pilot 2026.

UK Government Finds 407 Cyber Vulnerabilities Using Frontier AI

, , , , , , , , , , ,

The GC3’s month-long scanning program across nine departments uncovered critical flaws, including one that let any outsider run code on a government server by posting a comment.

 

The Government Cyber Coordination Centre (GC3), a partnership between the National Cyber Security Centre (NCSC) and the Department for Science, Innovation and Technology, published a case study on 12 June 2026 detailing a month-long series of weekly hackathons. Hacker teams used AI vulnerability scanning to review public code repositories across nine UK government organizations, working with frontier AI models including Claude Mythos and GPT-5.5. They identified 407 security flaws, including critical weaknesses that could let attackers bypass login controls, access sensitive data, or run code directly on government servers.

The most significant find: a flaw in a code repository used by a major government service that let any outsider run code directly on the government’s server by posting a comment. From there, an attacker could have used those credentials to approve code changes, reach connected systems, and remove any trace of the intrusion.

 

Teams built three distinct approaches.

  • One put each code repository through a six-stage AI pipeline, with each stage reviewing and pushing back on the conclusions of the one before it.
  • A second ran standard security scanning tools first to generate an initial list of potential issues, then used AI to trace how those issues could be chained into an actual attack.
  • A third built five custom AI tools that turned a complex audit across hundreds of services into something repeatable.

 

The GC3 drew four conclusions.

  1. Structure beats model choice. Teams that gave AI a defined job at each step consistently outperformed those using it without a clear structure.
  2. The specific AI model matters less than expected. With the right structure, slightly older models performed as well as the latest frontier ones at scanning code.
  3. AI finds potential issues faster than humans can check them. If the AI isn’t given clear boundaries on where to look, it flags far more potential issues than the security team has time to investigate.
  4. Finding vulnerabilities and fixing them are separate problems. All critical issues were patched through existing processes, but generating the actual fixes still requires human review.

 

GC3 will run a second phase with more departments, additional models, and coverage of government code that isn’t published publicly. AISI and NCSC will deepen involvement as the program moves further from controlled testing into live environments.

Netwrix 2026 Data and Identity Security Report cover, a geometric mosaic in dark navy and teal with the report title and subtitle.

Netwrix: Companies Deploying AI Breach At Four Times The Rate Of Peers

, , , , , , , , , , ,

Fast AI adopters breach at 43%, four times the 11% rate of slower peers, despite outperforming them on every security fundamental Netwrix measured.

Key Takeaways

  • Companies that have broadly deployed AI breach at 43%, four times the 11% rate for companies that have not
  • Fast AI adopters are ahead on security fundamentals, yet they breach more; AI deployments are creating new automated accounts and AI agents faster than security teams can track and control them
  • 76% of organizations do not fully track or control the AI agents and automated systems running inside their networks
  • 41% already have AI agents running in production, with those agents accessing company data on behalf of employees
  • 74% have no clear picture of which AI agents and accounts can access their sensitive data
  • Only 11% have AI governance that is fully enforced with consistent controls over who and what can access company data, applied in real time
  • Mid-size companies (500-999 employees) report the highest breach rate of any company size at 40%

Netwrix, a data security company, released its 2026 Data & Identity Security Report, finding that organizations in which AI has significantly increased the number of digital identities in their environments reported a 43% breach rate over the past 12 months. In security terms, an identity is any account with access to company systems: a human employee, an AI agent, a bot, or an automated service. Organizations where AI has not meaningfully expanded that number reported 11%. Netwrix surveyed 2,317 security and IT professionals across more than 60 industries worldwide.

 

Fast AI adopters are ahead on security fundamentals, yet they breach more

The gap is not explained by weaker security practices among faster AI adopters. Organizations leaning hardest into AI are measurably ahead on identity security fundamentals:

  • 46% lack a current picture of what data they hold and where it is stored, versus 60% among slower peers
  • 75% have strong controls over AI agents and automated systems with access to company data, versus 56% among slower peers
  • Only 13% cannot see which AI tools employees are using without company approval, versus 25% among slower peers

“They invested in the playbook. They got breached anyway,” wrote Grady Summers, Netwrix chief executive officer and a former Fortune 500 chief information security officer (CISO), in the report’s executive summary.

 

AI creates identities faster than governance tools can keep up

Summers says the problem is structural. Identity governance tools were built for environments where identities grew at the pace of hiring. AI agents, copilots, and automated services add new identities at the pace of software deployment. Every new agent is a new identity requesting access to sensitive data. Static or periodically reviewed permission models were not designed for that cadence.

The supporting data reflects that gap. 76% of respondents say they do not fully govern or monitor non-human identities in their environments. 41% say they already run AI agents in production, with those agents accessing company data on behalf of employees. 74% lack a single, unified view of sensitive data and which identities can reach it.

 

Mid-market breach risk is highest; only 11% have reached full governance maturity

The breach risk is highest in the mid-market. Organizations with 500 to 999 employees reported a 40% breach rate, the highest among company size ranges in the survey. Large enterprises with 10,000 or more employees reported 27.3%. Organizations with fewer than 500 employees reported 17.8%.

The report maps organizations across five AI governance maturity tiers. Just 11% have reached the level the report calls fully operational, meaning enforced, continuous, and proactive governance over every identity and the data it can reach.

 

With 41% of organizations already running AI agents in production and just 11% at full governance maturity, most are still building the controls that AI is already outpacing.

Abstract digital visualization of multiple translucent data panels with warning and alert icons flowing toward a glowing blue security shield with a checkmark, representing frontier AI's autonomous cyber threat capabilities

Frontier AI Can Now Execute Autonomous Cyberattacks. The Pace Is Accelerating.

, , , , , , , , , , , , ,

Key Takeaways

  • Frontier AI models can now autonomously execute cyberattacks against enterprise networks without human direction. Anthropic’s Mythos completed both of AISI’s simulated network attack exercises, including “Cooling Tower,” which no prior model had ever solved.
  • Autonomous cyber capability has doubled every 4.7 months since late 2024; Claude Mythos Preview and GPT-5.5 significantly exceeded that rate. 
  • Defenders and attackers have access to the same tools. Cyber defenders already report significant advances in vulnerability discovery using recent models.
  • AISI warns that cyber attackers will soon gain access to the same frontier AI models that are currently limited to a small pool of large tech companies, government organizations, and cyber experts. The time for organizations to invest in strong security is now.

AISI published new data confirming that frontier AI models can autonomously complete cyberattacks against enterprise networks. The rate at which those capabilities are improving is accelerating.

This is not a theoretical risk. AISI’s most recent testing found that Mythos successfully completed autonomous attacks against two simulated enterprise networks in the majority of attempts. One of those networks, “Cooling Tower,” had never been solved by any prior model.

AISI measures this progress using “time horizon” benchmarks: the time it takes a frontier model to reliably complete a task on its own, compared to how long the same task would take a human cyber expert. Since late 2024, that capability has been doubling every 4.7 months, itself an acceleration from the 8-month doubling time AISI reported in November 2025. Two recent models, Claude Mythos Preview and GPT-5.5, substantially exceeded even that faster trend.

Currently, access to Mythos and GPT-5.5 is gated. Anthropic granted access to only a few dozen organizations, including big tech companies, cybersecurity firms, and U.S. and allied government organizations. OpenAI limits access to GPT-5.5 through its Trusted Access For Cyber program, vetting users to ensure security. Over time, cybercriminals will find ways to access these tools and exploit cyber vulnerabilities. AISI stresses the time to implement strong security governance is now. 

AISI is developing tougher evaluation tools, including new cyber ranges with active defenses, to keep pace with model progress. The UK’s National Cyber Security Centre (NCSC) has published guidance on using AI models to find vulnerabilities.

Abstract visualization of interconnected software vulnerabilities and cyber threat detection linked to Anthropic’s Claude Mythos cybersecurity research.

Anthropic Says Its Claude Mythos Preview Identified More Than 10,000 Serious Software Vulnerabilities

, , , , , , , , , , , , , ,

Organizations participating in Anthropic’s Project Glasswing identified more than 10,000 serious software vulnerabilities, including flaws participants said previous methods likely would not have found. 

 

Anthropic released a detailed update on Project Glasswing, its restricted cybersecurity initiative that gives selected organizations access to Claude Mythos Preview (Mythos) to identify software vulnerabilities in widely used technology systems before attackers can exploit them.

Participating organizations identified more than 10,000 high- and critical-severity vulnerabilities during the first weeks of the program. The company said the vulnerabilities were found in “systemically important” open-source software projects heavily used across cloud infrastructure, operating systems, browsers, enterprise applications, and internet services.

Anthropic noted that several partners are finding more bugs faster, often by a factor of 10, outpacing human-guided efforts and even previous versions of Claude. Some examples include:

  • Cloudflare identified roughly 2,000 software flaws across critical systems, including 400 classified as high or critical severity. Cloudflare said the system produced fewer incorrect vulnerability alerts than its human security testers typically generate.
  • Mozilla said it found and fixed 271 vulnerabilities in Firefox 150 after testing with Mythos. Mozilla said that was more than ten times the number of vulnerabilities identified in Firefox 148 using Claude Opus 4.6, an earlier Anthropic model.
  • The UK’s AI Security Institute said Mythos became the first AI model to successfully complete both of its simulated cyberattack exercises, meaning the system was able to carry out an entire multi-step intrusion scenario from identifying a weakness to achieving the attack objective.

 

Finding open source threats

Anthropic used Mythos to scan over 1,000 open-source software projects that support major internet systems.

The company said Mythos identified 23,019 total vulnerabilities, of which 6,202 were estimated to be high- or critical-severity.

Anthropic said 1,752 of the high- or critical-rated findings have been reviewed by six independent security research firms, with a small number reviewed by Anthropic. Of the reviewed findings, 90.6% (1,587) were confirmed as real vulnerabilities. Anthropic said 62.4%, or 1,094, were confirmed as high- or critical-severity.

Based on those review results, Anthropic said Mythos Preview is on track to identify nearly 3,900 high- or critical-severity vulnerabilities in open-source code, even if it finds no additional flaws. The company said it plans to continue scanning open-source projects, so it expects that number to increase.

The company said most vulnerabilities identified through Project Glasswing have not been publicly disclosed because software maintainers are still developing fixes. Anthropic said vulnerabilities are handled through coordinated disclosure processes, in which affected organizations receive private notice before public disclosure.

 

Anthropic introduces Cyber Verification Program, countering OpenAI’s Trusted Access for Cyber (TAC) program

Anthropic also announced a separate gated cybersecurity program for security professionals. The program removes certain cyber guardrails, allowing security professionals already using Anthropic cyber models (Mythos excepted) for cybersecurity purposes to conduct real-world research and testing.

Anthropic said it plans to expand the number of Project Glasswing partners and increase collaboration with critical partners, including governments and infrastructure operators.