An artificial intelligence can now find and break into software largely on its own, and the company that built it decided that was dangerous enough to hold parts of it back. That should concern anyone in Malaysia responsible for keeping a network safe, because the country is trying to defend more of them than ever with roughly half the specialists it says it needs.

What you need to know:

  • OpenAI says its unreleased Astra model is the first it has rated "Critical" for cybersecurity, meaning it can find and exploit unknown flaws with limited human help.
  • In testing, Astra scored full marks on an exploit benchmark and uncovered two previously unknown vulnerabilities, which OpenAI says it reported to the affected software makers.
  • OpenAI is limiting who gets the strongest features, releasing them first to vetted testers.
  • Malaysia's Cyber Security Act 2024 now forces operators of critical systems to report incidents, while a government study puts the trained-workforce need at about 24,883 by 2026.

Table of Contents

What OpenAI actually said about Astra

On 1 September, OpenAI said Astra had crossed the "Critical" cybersecurity threshold in its Preparedness Framework, the first of its models to do so, according to the company and CNBC's coverage. With limited human guidance, the model can discover unknown software vulnerabilities, chain them into working exploits and run complex attacks. It scored 100 per cent on ExploitBench, a test of whether a model can build exploits for known flaws. Set against 20 high-severity bugs disclosed earlier in 2026, it identified and used two previously unknown "zero-day" flaws inside an exploit chain, and OpenAI says it disclosed those to the affected maintainers.

The company is not handing the full capability to everyone. It said it paused parts of Astra's development in August to add safeguards, and that the strongest cyber features will reach a small group of testers first before its Daybreak Blue security programme, as SecurityWeek reported. Fortune noted OpenAI plans to ship the model soon while withholding those functions from general users.

A technician inspecting network and server racks with a handheld diagnostic tool

Why this lands hard in Malaysia

Strip away the San Francisco framing and the question for a Malaysian bank, hospital or utility is simple. The hardest, slowest part of an attack, finding a fresh way in, is the part being automated. Defensive teams can use the same tools, and that is OpenAI's stated pitch. The problem is that Malaysia does not have enough people to run them.

A government workforce study projects the country will need about 24,883 trained cybersecurity specialists by 2026, up from a modelled demand of 2,460 in 2023, per the study's forecast. Prime Minister Anwar Ibrahim has put it in blunter terms, saying Malaysia needs around 25,000 trained specialists and has roughly 13,000. That is close to half.

Chart showing Malaysia needs about 24,883 cybersecurity specialists by 2026 but had roughly 13,000 in 2025, about half

This matters more now because the rules changed. The Cyber Security Act 2024 gave Malaysia its first standalone cyber law: a licensing regime for security providers and, crucially, mandatory incident reporting for operators of Critical National Information Infrastructure. These include the telcos, banks, power and water utilities, healthcare and government systems the economy runs on. NACSA, the national agency, can order assessments and coordinate the response when they are hit. A reporting duty is only as strong as the team behind it, and many of those teams are thin.

The asymmetry, and what to watch

None of this means attackers are running Astra tomorrow. OpenAI is limiting it precisely so they cannot, and the two flaws it found were handled responsibly. The uncomfortable part is that a capability, once demonstrated, tends to spread. Rivals build toward it, open models chase it, and the defensive comfort of "it is hard to find a new hole" erodes. Malaysia has bet heavily on becoming a regional data hub, and the same data-centre build-out that draws investment also widens the attack surface. Its own tug of war over digital rules shows the law can move quickly when it wants to. The workforce is the piece that cannot be legislated into existence overnight.

For Malaysian operators, the honest reading is not panic but priority. Automated offense is arriving faster than trained defenders, and the gap in the chart above is the one worth closing first.

Images courtesy of Kevin Horvat and Valentin Lacoste on Unsplash.