Good morning, security frontrunners. OpenAI has slowed Astra after preliminary testing suggested it may have crossed the company’s “Critical” cyber-capability threshold, while Google revealed an agentic discovery system that found more than 100 true-positive critical vulnerabilities in just two days.
Let’s get to it!
In this week’s Cyber AI breakdown:
OpenAI Slows Astra Development Over Possible “Critical” Cyber Capability
Harness Launches AI Agents Covering the Vulnerability Lifecycle
Atalanta Applies AI and Formal Methods to Satellite Cyber Resilience
US Army Cyber Trains AI Agents for Formal Cyber Work Roles
Google Details Its Agentic Vulnerability Discovery Harness
Anthropic Expands Claude Mythos 5 for Defensive Security
Latest Developments
The Breakdown: OpenAI has paused parts of Astra’s development after preliminary testing suggested the unreleased model may have reached its “Critical” cyber-capability threshold. The company is isolating cyber workloads and introducing token-level monitoring before allowing high-risk training and inference to resume.
The Details:
OpenAI paused deployment-focused reinforcement learning for two weeks and kept Astra’s largest planned frontier RL run on hold.
The company restricted code execution, tool use and internet access while removing shared services from high-risk training environments.
Monitoring classifiers now inspect every sampled token during Sol-and-above cyber evaluations, Astra training involving tools and Astra tool-enabled inference.
Automated investigators must examine serious alerts within 30 minutes; workloads pause if a potential critical-capability event cannot be disproved in that window.
OpenAI estimates that continuous monitoring consumes approximately 20% of inference compute, and several Astra cyber workloads remain suspended.
Why it Matters: Frontier-model development may increasingly be constrained by the models’ ability to autonomously find vulnerabilities, develop exploits and operate security tools, not merely by model-safety concerns. Cyber teams should expect powerful models to arrive with stricter access controls, extensive telemetry and compute-intensive monitoring requirements.
The Breakdown: Harness has introduced a collection of security agents designed to move vulnerabilities from initial detection through triage, remediation and deployment. The platform combines deterministic scanners with AI-generated fixes and compensating controls intended to protect systems while permanent patches are being prepared.
The Details:
The launch includes AI-powered SAST, scan orchestration, vulnerability triage, remediation, zero-day response and virtual-patching capabilities.
AI SAST combines deterministic code analysis with an AI noise filter and targets context-dependent flaws such as insecure direct-object references.
The Zero-Day Agent monitors new disclosures, maps affected packages to deployed pipelines and identifies which applications are genuinely exposed.
The Remediation Agent can generate and validate code changes before opening a pull request, while virtual patching can shield a vulnerable application without altering its source code.
Harness claims the integrated workflow can reduce vulnerability response from weeks to hours, although the performance figures have not been independently validated.
Why it Matters: The most valuable security agents may be those that connect discovery directly to validated remediation rather than simply producing more alerts. If the approach works at scale, AppSec teams could respond faster without allowing AI-generated vulnerability volume to overwhelm already constrained engineering resources.
The Breakdown: Atalanta has released Argo, a platform that combines AI-assisted software analysis with formal mathematical verification for high-assurance systems. Technology behind the platform was used to evaluate Viasat infrastructure following the destructive Russian-linked satellite attack in 2022.
The Details:
Argo analyses software and attempts to produce mathematically supported claims about how it will behave under defined operating and attack conditions.
Its underlying technology was applied to the Viasat satellite network affected by the 2022 KA-SAT cyberattack, which disrupted communications across Europe.
Atalanta describes Argo as the first commercially available platform of its kind, although that characterisation is a company claim.
The technology is also being used in the US Department of Energy’s Genesis Mission and research involving autonomous nuclear reactors.
DARPA has reportedly invested more than $2 billion over several decades in the formal-methods research underpinning this field.
Why it Matters: AI-assisted formal verification could provide stronger assurance than probabilistic code scanning for satellites, weapons, energy systems and other environments where failure carries physical consequences. It could also make techniques previously restricted to specialist research teams accessible to mainstream security and engineering organisations.
The Breakdown: US Army Cyber Command is training AI agents to perform formally defined roles alongside human cyber personnel. Rather than using a general assistant, the Army wants agents certified against the same job requirements used for developers, data engineers and cyber analysts.
The Details:
Initial roles include software developer, data engineer, host analyst and exploitation analyst.
Agents are being trained against the same Job Qualification Requirements used to certify human personnel for those positions.
A roughly 10-person organisation called Task Force Lexington is coordinating the Army’s agentic-AI work.
Current tasks include threat hunting, intrusion analysis, compliance documentation, eMASS and Risk Management Framework work, and agentic red-team operations.
The programme covers 17 cyber-protection-team mission elements, with humans retaining authority over risk-bearing and consequential decisions.
Why it Matters: Treating agents as defined cyber roles creates a more measurable operating model than simply giving analysts access to an AI chatbot. Commercial SOCs may eventually adopt similar certification, supervision and retraining frameworks for digital analysts working alongside people.
The Breakdown: Google Threat Intelligence has revealed the architecture of an agentic source-code analysis system it has operated for approximately ten months. The system combines specialised vulnerability-hunting agents, adversarial validation and human review to analyse very large stolen or exposed repositories during incident response.
The Details:
During one incident-response engagement, the harness reportedly identified more than 100 true-positive critical vulnerabilities in two days.
It has processed tens of millions of lines of code through thousands of pipelines, generating tens of thousands of candidate findings.
Dozens of results were considered assignable vulnerabilities, resulting in 12 issued CVEs and roughly another dozen active disclosures.
Google built the system using its Agent Development Kit, with deterministic orchestration controlling discovery, verification and reporting stages.
Dedicated validation agents are instructed to challenge findings before human specialists assess exploitability and disclosure requirements.
Why it Matters: Multi-agent review could make large-scale source-code analysis practical during breaches where defenders must rapidly determine what an attacker could exploit in stolen repositories. The sceptical validation layer is particularly important because unverified AI findings could otherwise bury incident responders in plausible but incorrect vulnerabilities.
The Breakdown: Anthropic is expanding access to Mythos 5 through Claude Security and selected defensive-security partners. The model will assist with vulnerability detection and remediation while keeping human review and narrowly scoped outputs around consequential findings.
The Details:
Mythos 5-powered code scanning is entering public beta for Claude Enterprise customers through Claude Security.
Reports include the suspected CWE, severity, confidence level, supporting analysis and a proposed remediation.
Human review remains mandatory, and Mythos 5 is not being exposed directly as an unrestricted offensive-security model.
Enterprise customers pay through standard token consumption rather than purchasing a separate security-scanning add-on.
Anthropic established a $35 million Defender Advantage Fund to support open-source patching, automated scan-and-fix projects and other defensive applications.
Why it Matters: Frontier-model providers are increasingly packaging advanced cyber capabilities into constrained defensive workflows rather than releasing unrestricted model access. For security teams, this could bring more capable code analysis while making evidence quality, permission boundaries and human approval central parts of deployment.
Everything else in Cyber AI this week
🏭 US agencies warned that attackers are using AI-generated Python tools to read and modify exposed Siemens PLCs across critical-infrastructure sectors.
🔓 Wiz’s Red Agent autonomously adapted an exploit and accessed a Snowflake Jira token through a vulnerable GitHub Actions workflow.
⚡ GPT‑5.6 Sol reportedly found 13 website-security weaknesses in 15 minutes and helped its operator remediate them in roughly one hour.
📈 Rapid7 found that high- and critical-severity vulnerability disclosures doubled year over year as AI accelerated vulnerability discovery.
📊 Beazley attributed a 36% quarterly surge in vulnerability disclosures partly to agentic AI, while compromised credentials still initiated 67% of ransomware deployments.
🦠 Fourteen trojanised npm packages delivered RedC2 4.0, whose AI-assisted Red Agent converts natural-language instructions into command-and-control actions.
🚦 Swimlane introduced intelligent SOC routing that assigns alerts to deterministic automation, AI assistance or autonomous investigation according to complexity.
🧩 Cribl acquired Radiant Security’s AI-SOC technology to generate alert-specific investigations directly against customers’ existing telemetry.
🤖 Cyberhaven says three autonomous security agents have reduced AppSec false positives, expanded threat modelling and resolved 70% of internal security tickets without human intervention.
That’s it for this week!
See you next Sunday 🙂
Zac S from The Cyber Breakdown