← The Wire·Breaking·Cloud··3 outlets

OpenAI says its new 'Astra' AI can build attacks without human help

OpenAI says its new 'Astra' model is the first to hit its 'Critical' cybersecurity threshold, capable of discovering unknown vulnerabilities and building exploits against hardened systems without human help. This raises the stakes for both offensive AI risk and the demand for AI-driven cyber defense tools that can counter it.

The story

OpenAI disclosed that its newest model, internally named Astra, has crossed the company's own 'Critical' threshold on its cybersecurity capability framework — the first model to do so. This means Astra can independently identify previously unknown vulnerabilities (zero-days) in software and hardware, and then develop working exploits against systems that have already been hardened against known attack techniques, without requiring a human operator to guide the process.

OpenAI has published internal risk thresholds for its models for some time, categorizing capabilities across domains like biological, chemical, cyber, and persuasion risk as part of its preparedness framework. Previous generations of models were assessed as falling short of the 'Critical' bar in cybersecurity, meaning they could assist human attackers but not autonomously complete the full vulnerability-discovery-to-exploit chain. Astra's classification marks a step change in the company's own internal risk taxonomy, not a third-party or regulatory assessment.

The disclosure touches multiple constituencies. OpenAI itself faces scrutiny over how it plans to deploy or restrict Astra given the classification, including whether it will be released broadly, gated behind enterprise agreements, or held back from public access. Cybersecurity vendors — companies building AI-assisted defense, vulnerability scanning, and threat detection — stand to benefit from heightened urgency among enterprises and governments to adopt AI-augmented defenses capable of matching AI-augmented offense. Cloud providers hosting AI workloads and enterprises running legacy or hardened infrastructure are the parties most exposed if such a model's capabilities proliferate or are replicated by less careful actors.

The classification is based on OpenAI's own internal safety evaluation framework and has not been independently verified by outside researchers or regulators at this stage. It remains unclear how OpenAI intends to mitigate the risk in practice — whether through access restrictions, monitoring, or technical safeguards — and what specific evidence supports the claim of full autonomy in exploit development.

Watch for how OpenAI structures access to Astra, any accompanying safety card or technical report detailing the specific vulnerabilities or systems tested, and reactions from cybersecurity researchers or government bodies that may seek to validate or challenge the 'Critical' designation. Any regulatory response, particularly from bodies focused on AI safety or critical infrastructure protection, would be a key marker of how seriously this disclosure is treated.

Our read · Fri, Sep 4

No single public equity is directly named, but the disclosure sharpens the case for AI-driven cybersecurity vendors (e.g., CrowdStrike, Palo Alto Networks, Fortinet) as enterprises face pressure to counter autonomous exploit-generation capability.

Why. OpenAI is privately held; the read-through is thematic — heightened autonomous-exploit risk should, in theory, increase enterprise urgency around AI-augmented cyber defense spend, but that flow-through is diffuse and unquantified in the reporting.

Both cases

The case it holdsIf Astra's capability is real and disclosed publicly, enterprises and governments may accelerate purchases of AI-native cybersecurity and vulnerability-management tools to keep pace, a tailwind for defense-focused vendors.
The case it breaksThe classification comes solely from OpenAI's internal framework with no independent audit, named vulnerabilities, or third-party confirmation, so the claim could overstate real-world exploit risk and the anticipated defense-spend response.

What could change this view: The classification is OpenAI's own internal self-assessment with no independent verification, no named systems tested, and no public technical detail, so the actual real-world risk and any resulting defense-spend uplift are both unconfirmed.

Your side is graded privately against closes after 10 trading days. Research, not advice.

Who reported it · 3

Same arc

Reports linked to this one, newest first. Automatically linked reports may cover a broader event.

OpenAI Holds Back Astra Over Safety Risks, Iran Doubts US Deal Before MidtermsBloomberg Television · Tue, Sep 29 · 6:25 AM ET · automatically linked
OpenAI says governments among ‘dozens’ of organisations hacked by its agentsFinancial Times · Fri, Sep 25 · 6:18 PM ET · automatically linked
OpenAI expects to burn $280bn by 2030Financial Times · Fri, Sep 18 · 6:33 PM ET · automatically linked
Why Oracle Stock Jumped 6% Today on OpenAI Funding BuzzYahoo Finance · Thu, Sep 17 · 1:16 PM ET · automatically linked

Keep reading

Latest on the Wire

Russian cargo ship Sparta II reportedly hit by drones near FinlandBTCS warns $54 million Aave loan faces liquidation as ether fallsIntesa Sanpaolo launches €34.5bn cash-and-share offer for Monte dei PaschiIran war sees tanker hit near Hormuz as Saudi Arabia intercepts Houthi missilesNvidia tumbles as US tech stocks enter bear market on AI spending fearsContinue on the Wire →

Top this week

Nvidia allies lobby board to bring back Jensen Huang as CEONvidia chair rules out Huang return as CEO after 75% sales outlookAnthropic Said to Guide IPO Pricing Above Market RangeNvidia interim CEO Colette Kress faces first earnings test after HuangAnthropic IPO could become the largest in US history on Thursday
The Wire at 7 AM ET, free. Top stories with a bull and a bear case on each, every morning. No account needed.

About this report

Reported by CoinDesk, . Who answers for this

3 reports from 2 publishers in this record. Closes are licensed end-of-day data from Databento; the move after a story is its first close against the close before it. How this is decided →

Kept as written · your side, if you take one, is graded privately against licensed closes after 10 trading days · nothing here is advice · How the Wire is made →