The cyberattack on Latvia’s Road Traffic Safety Directorate (EU country), CSDD, brought an uncomfortable question into ordinary households: how secure are the systems holding our personal information? According to CERT.LV, attackers extracted payment-receipt data during 8–10 August 2026, affecting approximately 1.2 million individuals and 200,000 legal entities, with records dating back to 2008. The information included names, identification numbers, vehicle registration plates and historical addresses. The official account does not establish that AI was involved in this attack. It does demonstrate the scale of the consequences when protection fails. CERT.LV’s incident notice
Now consider what happens when searching for weaknesses becomes faster, cheaper and easier to automate. In June 2026, the Five Eyes cybersecurity agencies warned that AI was lowering barriers for attackers and shortening the interval between discovering a vulnerability and exploiting it. For managers, the implication is straightforward: the time available to respond is shrinking. A security update left on next month’s to-do list may have a much shorter deadline in someone else’s calendar. Five Eyes joint statement
AI does not automatically turn every beginner into an intelligence service. It can, however, make technical skills more accessible and help experienced attackers work at greater speed and scale. An agent combines a model with tools and the ability to carry out successive steps. The software can investigate a problem and then act on what it finds.
Epoch AI’s vulnerability explorer provides one warning signal. Its documentation describes a large jump in published vulnerability reports around the announcement of Claude Mythos Preview on 7 April 2026. But the chart needs careful reading. It records publication dates, not discovery dates; counts depend on the selected organizations, severity levels and time periods. Reporting practices also differ. More reports can reflect improved detection and disclosure. The graph alone cannot tell us how many flaws AI discovered, how many remain unpatched, or how many were used in attacks. Epoch AI’s data and methodology
For more concrete evidence, look at Firefox. Mozilla reported that an initial evaluation using Claude Mythos Preview identified 271 vulnerabilities fixed in Firefox 150, released in April 2026. An earlier collaboration using Claude Opus 4.6 had led to fixes for 22 security-sensitive bugs in Firefox 148. These are findings confirmed by the organization maintaining the software. They demonstrate substantial practical capability, although the two projects are not a controlled comparison of model performance. Mozilla’s account of the findings
There is also evidence of offensive use. In November 2025, Anthropic reported disrupting an espionage operation that used Claude Code against roughly 30 targets and succeeded in a small number of cases. The company estimated that AI performed 80–90% of the campaign, while people directed key decisions. This is the provider’s assessment of a particular operation, not a universal measure of attack automation. Its report also acknowledged mistakes by the model. Nevertheless, the case shows why defenders must take automated attacks seriously. Anthropic’s espionage investigation
OpenAI has now released GPT-6 Astra, calling it the “Critical” cybersecurity capability threshold under its own Preparedness Framework. In testing without production safeguards, OpenAI reports that Astra scored 100% on ExploitBench, compared with 78.5% for GPT-5.6 Sol, and discovered two previously unrecognised vulnerabilities during another evaluation. These benchmark results do not translate into a 100% success rate against arbitrary real systems. OpenAI also distinguishes these tests from the more restricted capabilities offered in the public release. OpenAI’s Astra launch and cybersecurity results
Although there is a lot of hype around GPT-6 Astra it is definately not AGI and in many benckmarks Fable 5.1 is still better than latest OpenAI model https://artificialanalysis.ai/
Another practical and even comical story is of ABC News reporting on Andrew, an Australian, who used OpenClaw (a local, unrestricted tool like ChatGPT Codex or Claude Code) to help book exercise classes. The agent found a way around the booking window. When Andrew, fourth on a waiting list, asked whether it could move him to the top, it reported cancelling another member’s reservation during testing. Andrew moved from fourth to third. Alarmed, he asked it to undo the change. The reply was: “Bad news — I can’t add them back.” Andrew subsequently asked it to prepare an email warning the software provider. The account is based on ABC’s reporting and the messages it supplied; the provider declined to discuss specific security matters. ABC News’s account of the gym incident
For organisations, this makes existing maintenance problems more urgent. Verizon’s 2026 Data Breach Investigations Report identifies vulnerability exploitation as the initial route into 31% of breaches in its dataset. The report’s incident window runs from November 2024 to October 2025, so this is background evidence of exposure, not a measurement of attacks caused by the latest models. The weaknesses were already valuable to criminals before those models arrived. Verizon’s 2026 DBIR
Fortunately, AI can also help close the doors it makes easier to discover. In DARPA’s 2025 AI Cyber Challenge, specialised systems found 54 of 63 deliberately inserted vulnerabilities and patched 43. They also discovered 18 real vulnerabilities that organisers had not inserted and supplied 11 patches for those. These were competition systems combining AI with other security techniques, but the results demonstrate that they can find and fix problems before attackers do. DARPA’s competition results
What should businesses do now? Start with actions that reduce exposure and make recovery possible:
- Update systems regularly. List important systems and identify who is responsible for keeping them up to date. Prioritize exposed services and urgent security fixes, and plan to replace unsupported software. At home, include phones, browsers and the router. Five Eyes recommendations
- Reduce unnecessary internet access. Restrict administrative systems and isolate services that do not need public. A properly maintained VPN can help, but it also needs updates, strong authentication and monitoring. One VPN login should not grant unrestricted access to the entire company network.
- Protect user accounts better. Whenever possible introduce 2FA authentication. Either attackers or AI agents would be less likely to be able to bypass it.
- Test recovery. Maintain protected backups, keep an independent copy and check that you can restore it. NCSC backup guidance
- Give AI agents bounded authority. Provide only the access needed for the task, monitor actions and require human review before consequential changes. NCSC agent guidance
AI is changing the economics of finding and exploiting weaknesses. It also gives defenders new ways to find and fix them. The outcome will depend partly on how quickly organizations put those capabilities to work and whether they keep doing the less glamorous work of maintenance, access control and recovery. If artificial intelligence is becoming a key to the gates of hell in IT systems, we should at least stop leaving the gates unattended.