OpenAI Astra Sets a New Standard for AI Cyberattacks
According to OpenAI, Astra can find zero-days on its own and build exploits. For wallets, exchanges, and smart contracts, that raises the risk of lightning-fast attacks.
Key Takeaways
- OpenAI says the upcoming Astra model can independently find software flaws and carry out attacks without human direction.
- Astra received the cybersecurity status “Critical” within OpenAI's Preparedness Framework and scored 100% on an exploit benchmark.
- For crypto, this is especially relevant because vulnerabilities in wallets, exchanges, smart contracts, and infrastructure can quickly turn into money.
OpenAI says its upcoming Astra model can find software flaws and turn them into working attacks on its own, without a person guiding every step. According to the company, Astra is the first model to receive the cybersecurity status “Critical” within its Preparedness Framework. For the crypto market, that matters because a software flaw can be turned into money in a short time.
What Astra Can Do
OpenAI writes that a model must be able to find unknown flaws, also known as zero-days, for this status and build working exploits for them in hardened, real-world systems. In tests, Astra scored 100% on a benchmark for developing exploits based on known vulnerabilities. It also found two previously unknown flaws while working on an exploit chain in an internal test.
The model also managed to break out of a hardened browser sandbox and run commands on the host computer. In another test, it combined multiple flaws in an operating system to gain root access, according to OpenAI. That shows how far advanced AI can now go, not just in writing code but also in carrying out attacks step by step.
Why Crypto Is Extra Vulnerable
For crypto, this is especially sensitive because attacks on wallets, exchanges, smart contracts, or infrastructure often pay off quickly once a weak spot is found. Security researchers previously warned that AI can sharply shorten the time between finding code flaws and carrying out an attack. That does not mean every attack will automatically succeed, but it does mean the speed at which weak spots are exploited could keep rising.
That development fits into a broader trend where frontier models are going beyond answering questions or writing code. Other AI companies are also showing more often that their models can handle complex technical tasks, which increases the pressure on cybersecurity teams.
OpenAI Slows the Rollout
OpenAI has now delayed parts of Astra's development while it adds extra safety measures. The company wants to make the most advanced cybersecurity features available only to selected testers at first. In doing so, OpenAI is acknowledging that the model's capabilities are large enough to make misuse a serious concern, even before it is rolled out more broadly.
The concerns line up with broader warnings about fast AI-driven cyber threats. In a recent analysis from the Five Eyes, it was already stated that frontier AI can sharply shorten the time between finding a vulnerability and exploiting it.
