Good morning, security frontrunners.
In today’s Cyber AI breakdown:
GPT‑5.6 Sol Raises the Bar for General-Purpose Cyber Capability
Alberta Uses 50 Claude Code Agents to Audit 466 Million Lines of Government Code
EU Launches a Bloc-Wide Plan for AI-Powered Cyber Defence
AI Red Agent Reaches an Airline’s Booking Database in 15 Minutes
JADEPUFFER Signals the Arrival of Autonomous Ransomware
Latest Developments
The Breakdown: OpenAI has released GPT‑5.6 Sol globally, positioning it as its strongest model yet for vulnerability research, exploitation and other long-horizon cybersecurity tasks.
The Details:
GPT‑5.6 launched on 9 July across ChatGPT, Codex and the OpenAI API, with Sol as the flagship model and Terra and Luna providing lower-cost options.
Sol scored 96.7% on OpenAI’s capture-the-flag challenges, 71.2% on SEC-Bench Pro, 73.5% on ExploitBench and 33.7% on ExploitGym. Results comparable to Mythos Preview but not quite as good as Mythos 5.
Sol’s Ultra setting coordinates multiple subagents across parallel workstreams and increased its SEC-Bench Pro result to 74.3%.
Sol is priced at $5 per million input tokens and $30 per million output tokens, which makes it significantly cheaper than Mythos 5 at roughly $10 and $50 respectively.
OpenAI says GPT‑5.6 did not cross its “Critical” cyber-capability threshold and cannot yet reliably complete autonomous attacks against hardened targets.
Its safeguards reportedly block around 10 times more potentially harmful activity than earlier models, following approximately 700,000 A100-equivalent GPU hours of automated red-teaming.
Why it Matters: High-end vulnerability research and exploit analysis are becoming accessible through general-purpose models rather than specialist security tools alone. Cyber teams must determine how to use these capabilities defensively while preparing for faster vulnerability discovery, proof-of-concept development and attack automation by adversaries.
The Breakdown: Alberta’s Ministry of Technology and Innovation used approximately 50 parallel Claude Code agents to review 466 million lines of code across the provincial government, completing the initial scan in around 20 hours.
The Details:
The reviewed estate supports all 27 provincial ministries and comprises approximately 1,280 applications across 3,400 code repositories, much of which had never received a systematic security review.
Around 50 autonomous agents scanned 466 million lines in 20 hours; Alberta estimates an equivalent manual review could have taken approximately 6.5 years.
The scan used a two-stage pipeline: a rules engine first identified known vulnerability patterns, after which Claude analyzed the candidates and cited the exact affected files and lines for developer verification.
Claude could generate, compile and test fixes automatically; where applications lacked sufficient test coverage, it first created the tests required to validate that a patch would not introduce regressions.
Alberta also created dedicated red and blue team agents. The red agent models external exploitation paths, while the blue agent evaluates applications against approximately 95 security controls and produces file-level remediation plans.
Why it Matters: The case study demonstrates how agentic code analysis can turn vulnerability discovery from a periodic, sample-based exercise into an estate-wide process spanning hundreds of millions of lines of legacy code.
The Breakdown: The European Commission has launched a coordinated plan to expand the EU’s use of AI for vulnerability detection, incident prevention and critical-infrastructure protection. It combines defensive deployment, model evaluation and investment in sovereign European cyber-AI capabilities.
The Details:
The plan has three objectives: promote responsible advanced-AI use, reinforce EU cyber resilience and scale Europe’s AI capabilities for cybersecurity.
ENISA and the Joint Research Centre will develop a secure platform where critical sectors, including energy, healthcare, transport, finance and government, can test cyber-AI systems.
Organizations will be encouraged to use AI, including appropriate open-source models, to discover and address vulnerabilities more rapidly.
A new EU Grand Challenge will bring together researchers and companies to develop innovative AI-powered cybersecurity solutions.
The plan links deployment to NIS2, the Cyber Resilience Act, DORA and the Cyber Solidarity Act while supporting sovereign capacity through AI Factories and future Gigafactories.
Why it Matters: Cyber AI is becoming part of national and regional security infrastructure rather than remaining a collection of commercial products. The plan could influence procurement, testing standards and regulatory expectations for every security vendor or enterprise operating in Europe.
The Breakdown: Wiz says its autonomous Red Agent started with only an unnamed airline’s public homepage and reached its booking database within 15 minutes. It discovered a conventional object-level authorization failure, but did so by reasoning across JavaScript, authentication flows and a GraphQL schema rather than matching a known signature.
The Details:
The agent received no credentials, seeds or internal information and began with a single public root URL.
It analyzed client-side JavaScript, identified the API gateway and replayed a multi-stage token flow to mint a valid anonymous session.
GraphQL introspection exposed 514 queries and 428 mutations to that anonymous session.
The agent tested 20 sequential booking IDs and received a different customer record every time, exposing approximately two years of passenger and itinerary data.
The flaw provided read and write capabilities, including access to names, birth dates, contact details, billing addresses and masked cards, plus functions capable of cancelling flights, changing contacts, overriding prices and issuing refunds.
Why it Matters: The demonstration shows that autonomous testing can uncover business-logic and authorization flaws that conventional scanners rarely detect. Defenders should expect attackers to automate this reasoning, making resolver-level authorization, restricted production introspection and non-predictable identifiers increasingly urgent controls.
The Breakdown: Sysdig assesses JADEPUFFER as the first documented ransomware operation driven end-to-end by an AI agent, covering reconnaissance, credential theft, lateral movement, persistence and destruction. Its ability to diagnose failures and immediately alter its approach distinguishes it from conventional automation or a fixed attack script.
The Details:
Initial access came through CVE‑2025‑3248, an unauthenticated remote-code-execution flaw in an internet-facing Langflow deployment.
The operation moved across two targets: the compromised Langflow host and a separate production server running MySQL and Alibaba Nacos.
Sysdig captured more than 600 purposeful payloads; following one failed login, the agent diagnosed and corrected the problem in just 31 seconds.
JADEPUFFER searched for LLM API keys, AWS, Azure, GCP and Chinese cloud credentials, cryptocurrency wallets, database secrets and configuration files, while installing a beacon that ran every 30 minutes.
It encrypted all 1,342 Nacos configuration records and dropped the original tables, but failed to preserve or transmit the encryption key, meaning payment could not restore the victim’s data.
Why it Matters: Autonomous agents could dramatically reduce the expertise and staffing required to run an end-to-end ransomware operation. Defenders must prepare for attacks that branch, retry and adapt at machine speed, particularly around exposed management platforms, credential-rich AI infrastructure and long-unpatched vulnerabilities.
Everything else in Cyber AI this week
Sygnia’s 72-hour AWS investigation shows how one apparently AI-assisted attacker compressed weeks of cloud discovery, credential theft and extortion activity into three days.
The UK’s proposed Cyber Shield would use federated red and blue agents to identify vulnerabilities and eventually automate national-scale remediation.
CISA is reportedly using Anthropic’s Mythos to search US government code repositories for vulnerabilities exploitable by foreign spies and cybercriminals.
ESET’s latest threat report highlights generative AI operating inside Android malware and thousands of malicious skills within the emerging agent-tool ecosystem.
Blackpoint’s new identity-focused AI SOC agent claims it can autonomously contain certain Microsoft 365 and Google Workspace attacks in as little as 21 seconds.
Reach Security’s Network Security Assurance applies AI to continuously identify and remediate firewall, SASE and network-control drift.
Picus introduced an AI-agent swarm designed to combine penetration testing, breach simulation and fix validation into one continuous defensive loop.
That’s it for this week!
See you next Sunday 🙂
Zac S from The Cyber Breakdown