Astra 6 learned to find zero-day vulnerabilities

AI has crossed a new line. Astra 6 can already identify previously unknown zero-day vulnerabilities, validate them, and build working exploit chains. For the crypto market, that is a serious warning: a huge share of blockchain infrastructure, wallets, DeFi protocols, and smart contracts runs on open-source code that directly controls real money. That code can now be probed for weaknesses continuously.

OpenAI says Astra has reached a critical level of cyber capability. During testing, the model did not simply reproduce known attack techniques. It discovered two previously unknown zero-day vulnerabilities and was able to use the weaknesses it found as part of real exploit chains.

This is no longer just an AI coding assistant. It is a system capable of identifying a weak point, testing it, and understanding how that flaw can be turned into a working attack.

What Astra 6 was able to do

In one test, the model worked against a hardened browser and operating system. Astra identified vulnerabilities, escaped the browser sandbox, and gained the ability to execute commands on the host system. In another scenario, it combined several flaws and escalated privileges from a normal user account to root.

On ExploitBench, where models are tested on their ability to create exploits for known vulnerabilities, Astra achieved a 100% result. OpenAI then tested the model against a separate set of 20 recently disclosed high-severity vulnerabilities. During those evaluations, the system discovered two additional real vulnerabilities that had not previously been known to the software maintainers.

That is the key difference from traditional AI tools used by developers. The model is not merely helping a human analyze code. It can identify a weakness, verify it, and determine how that weakness can be exploited.

Why this matters directly to crypto

A large part of the cryptocurrency ecosystem is open to public inspection. Bitcoin Core is open source, as are Ethereum clients, Lightning implementations, many wallets, bridges, libraries, and most DeFi smart contracts.

For years, that openness has been treated as one of blockchain’s core strengths. Any developer or security researcher can review the code, identify a flaw, and report it before it causes damage. But the exact same code is also available to an attacker, and now to AI agents capable of reviewing it at a scale far beyond manual analysis.

The main change is not that vulnerabilities suddenly exist. They always did. What changes is the cost and speed of finding them.

More than $100 billion is already sitting inside open code

OpenAI has already been testing AI specifically against blockchain software. Together with Paradigm, it introduced EVMbench, a benchmark designed to measure how well AI agents can discover, patch, and exploit smart contract vulnerabilities.

EVMbench includes 117 real vulnerabilities drawn from 40 audits. The reason for building such a benchmark is obvious: open smart contracts routinely control more than $100 billion in crypto assets.

That creates an unusually attractive environment for automated vulnerability discovery. The code is visible, the change history is public, dependencies can be inspected, and contract balances are often visible directly on-chain. If an AI system finds a critical flaw, it may not need to search long for a way to monetize it. The money can already be sitting inside the system being analyzed.

The real threat is scale

In the past, a complex audit could take weeks. A security researcher had to understand a project’s architecture, inspect dependencies, analyze thousands of lines of code, build hypotheses, and reproduce a potential vulnerability.

AI changes that model completely. One specialist can run multiple agents at the same time, while those agents inspect dozens or hundreds of repositories, new software versions, library updates, and smart contracts in parallel.

Most of those checks will find nothing critical. But an attacker does not need a successful result every day. One exploitable flaw in a system holding tens or hundreds of millions of dollars can be enough.

The crypto industry is already reacting

In August, more than 40 crypto companies and organizations called on major AI labs to give Bitcoin security researchers and other open-source defenders access to their most capable models.

Signatories included Coinbase, Block, BitGo, Blockstream, ARK Invest, Exodus, Foundry, and Casa.

Their concern is simple. Defenders of open-source infrastructure may be forced to work with restricted public AI models, while attackers can use local, modified, or specialized systems with fewer safeguards.

For crypto, that imbalance is particularly dangerous. The code behind Bitcoin, Lightning, and many other systems is already public. An attacker does not need to break into a company first to steal source code. The target is already available for analysis.

OpenAI is putting $1 billion into defense

Against that backdrop, OpenAI launched Daybreak for Frontline Defenders, a program backed by $1 billion in subsidized access to AI models, technical support, training, and defensive tools.

Priority areas include critical infrastructure, government organizations, banks, and open-source projects. The last category matters directly to the crypto industry because much of Bitcoin, Ethereum, and the wider blockchain ecosystem is built on open software.

This effectively creates a new technological arms race. One AI system searches for a vulnerability so it can be fixed first. Another may be looking for the same flaw for the opposite reason. The advantage goes to whoever finds the weakness first.

The old audit model is becoming weaker

Until now, a crypto project could pass one audit before launch, publish the report, and use the auditor’s name as proof of security. But that audit only captures one version of the code at one point in time.

After the review, libraries are updated, contracts are changed, integrations are added, and dependencies evolve. If the attacking side can analyze a system continuously, a one-time audit quickly becomes less meaningful.

The next security standard is likely to be continuous AI auditing. Every new commit, contract change, or library update can be tested immediately before it reaches users. Without that, defenders risk losing the race on speed alone.

The blockchain itself does not need to be broken

Stealing crypto does not require breaking the mathematics behind Bitcoin or Ethereum. It is often enough to find a flaw in one of the many software layers around them.

That could be a wallet, browser extension, bridge, Lightning node, exchange API, update server, library, or smart contract. The blockchain can continue operating perfectly while funds disappear because of a flaw in ordinary software.

That is why AI systems capable of finding zero-day vulnerabilities matter so much to the crypto market. In this industry, code is often directly connected to money.

Kljompus Conclusion

Crypto has spent years treating open-source code as a strength. Now that same openness is becoming a new attack surface.

In the past, people reviewed the code. Now AI agents can inspect it around the clock across hundreds of projects at once.

The next major race in crypto may not be about blockchain speed, liquidity, or a new token. It may simply be about who finds the bug first — the defender or the AI that came for the money.

Sources

OpenAI — Path to Astra: Critical Capabilities and Frontier Safeguards

OpenAI — Safety Overview: GPT-6 Astra

OpenAI — Daybreak for Frontline Defenders