Security | Threat Detection | Cyberattacks | DevSecOps | Compliance

For the ten years I have watched this industry, hunting business logic vulnerabilities has been notoriously expensive and exceedingly difficult. Every attempt to bring the cost down has hit the same wall. Now, for the first time, we as an industry have the capability to meaningfully lower the ceiling on what this work costs. I would argue Astra’s autonomous pentesting platform is pioneering that shift.

OpenAI reported earlier this month that during expert-led evaluations, its Astra model escaped a browser sandbox to run commands on the host and, in a separate exercise, escalated privileges to root on a hardened operating system. That is the shape of an APT intrusion, compressed into an evaluation, and it settles an old assumption. Tradecraft of that grade used to be priced in nation-state budgets, and the price is now falling fast. Our concern at Astra Security is what that means for how companies test.

I’ve watched AI-powered cyberattacks inch closer to autonomy throughout 2026, and I can point to dozens of real-world incidents as proof. AI-enabled cyberattacks surged by 56% over the last year, meaning 1 in 4 (25%) malicious breaches now involve artificial intelligence. The important one: on May 10, 2026, researchers documented the first fully autonomous LLM-driven post-exploitation attack in the wild.

Change Healthcare took down a third of US claims processing. Ascension spent weeks on the papers. OCR settlements keep citing “risk analysis failure,” and HITRUST r2 assessors are asking harder questions about what actually got tested versus what got scanned. The math on pentesting shifted in the middle of all that. In 2024, 1 in 40 findings was Critical. In 2025, it’s 1 in 10.

On September 1, 2026, OpenAI announced that its Astra model had reached the Critical tier for cyber capability under the company’s Preparedness Framework, a first for its systems. While working through an internal benchmark of 20 recently disclosed vulnerabilities in Google’s V8 engine, the model found two zero-days that no one had asked it to look for and used them in a working exploit chain. OpenAI is disclosing both flaws and restricting the capability to vetted testers.

Security leaders at key financial institutions in Singapore now have a new line in their compliance calendars: AI-assisted red teaming, a requirement MAS introduced on 1 July 2026. What appears to be a scoping exercise actually asks a harder question, i. e., how does a bank differentiate between another convincing report and one that secures the institution?

On September 1, OpenAI announced that its new model, GPT-6 Astra, had become the first to cross the “Critical” cybersecurity threshold in the company’s Preparedness Framework. Most coverage focused on the safety implications, and fairly so, but buried in the announcement sits a benchmark result that should change how every security leader reads their next vulnerability report.

Buried in OpenAI’s evaluations of its GPT-6 Astra model sits a finding that mobile teams should sit with for a minute. The model reverse-engineered compiled software well enough to escape a browser sandbox and chain privilege-escalation flaws on a hardened operating system, according to the company’s published evaluations. Reading compiled binaries used to be specialist work priced in weeks, and now it’s something machines do quickly and well.

Every security vendor’s homepage now says “AI-powered” somewhere above the fold, and most are describing the same scanners they sold in 2022 with a model bolted onto the reporting layer. Buyers have noticed, and the skepticism is earned. When everything claims to be intelligent, the label stops conveying information. A security lead evaluating tools is left with one question that matters: whether the tool can actually think through an attack the way a pentester does.

The security industry killed CVE coverage as a credible metric, and it deserved to die. Vendors inflated the numbers for years in the name of depth, and nobody in the room had an incentive to ask whether they reflected real validated risk or just a longer signature list. So “CVE coverage is a vanity metric” became earned consensus. The question I keep coming back to is whether the autonomous pentesting era makes that consensus outdated.