OpenAI unleashed GPT-6 Cyber to crack Linux and pentest software
Automated hackers are officially here, and this time they aren't rogue scripts from dark web forums. OpenAI built a dedicated digital intruder that tears through operating systems and web apps without needing to see a single line of source code.
The new security-focused heavyweight rolled out under the name GPT-6 Cyber, built specifically to automate white-hat defense and nasty offensive penetration testing. To prove its chops right out of the gate, engineers unleashed the model on Linux and the Firefox browser. The audit unearthed 948 ways to escalate system privileges in Linux and flagged 81 known security flaws in Firefox, proving that even rock-solid open-source giants leak like rusted buckets when machine learning pokes the code.
Instead of relying on public documentation, the model thrives on black-box analysis to rip through compiled binaries and live web infrastructure completely blind. To contain the chaos, OpenAI split operations into two dedicated tracks within its Daybreak program: Daybreak Blue focuses on patch creation and vulnerability triage, while Daybreak Red plays the adversarial villain, engineering live exploits. Guardrails are supposedly hardcoded to keep the neural network within defined IP scopes so it doesn't spontaneously decide to colonize neighboring servers like an autonomous digital plague.
Benchmark numbers show clear algorithmic escalation across offensive categories. On the ExploitGym benchmark, which measures turning known vulnerabilities into functional weaponized exploits, GPT-6 Cyber hit a 48.5% success rate, stomping the 32.1% score managed by Cyber-5.6. On raw web pentesting without code access, the system snagged 40.3% of target vulnerabilities against 32% achieved by GPT-6 Astra, while crushing the SRE-Bench reverse-engineering test with an eye-popping 92.8% accuracy. Access remains locked down to vetted Daybreak partners.
Corporate gatekeeping will only hold until autonomous attack models become standard commodity software. When high-schoolers with rented API keys can run automated zero-day campaigns faster than enterprise security teams can brew morning coffee, the entire bug bounty industry morphs into an algorithmic eviction notice for human white-hats.
Source: OpenAI
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.