AI is finding flaws faster than defenders can patch them, while autonomous agents push past human instructions. From exposed government systems to an Anthropic model that ignored stop signs, the question is who can halt the machine—and who pays when it acts without permission or moves the money.
Thursday's AI sell-off was not about demand. It was about who pays. The Nasdaq fell 2.6 per cent, Meta unveiled non-voting shares to fund its build-out and Nvidia's Jensen Huang warned of a scarcity of capital, as OpenAI raced to close up to $30 billion at $1.25 trillion. Retracement or reckoning?
SpaceX posted a US$4.4 billion quarterly profit as AI revenue surged, while Tesla and SpaceX courted Korean suppliers for their Terafab chip project in Texas. Musk is targeting production in 2028 as the companies also explore a potential merger.
AI is finding flaws faster than defenders can patch them, while autonomous agents push past human instructions. From exposed government systems to an Anthropic model that ignored stop signs, the question is who can halt the machine—and who pays when it acts without permission or moves the money.
Within a day, security teams had named the week Black October. The two PHP2Shell flaws, hidden for more than a decade in the language behind roughly 70 per cent of the web, were found with machines: by Calif using OpenAI’s GPT-5.6 Sol and, independently, by Wiz with Google’s Gemini 4 Argon.
OpenAI turned the find into Patch the Planet, a $1 billion pledge of subsidised access for stretched defenders and a showcase of AI’s gift to whoever finds a flaw first: time.
The head start ended the moment the secret did. Details leaked a day early, European governments pulled websites offline and the Australian Signals Directorate issued its highest alert.
By the weekend Palo Alto Networks had found attackers inside a European government body, holding password hashes and multi-factor secrets within two hours, their commands formatted for an AI to read. A campaign tentatively tied to North Korea’s Lazarus ecosystem has listed 216 targets from banks to airports, Australia among them.
“Every single week it’s going to feel like the sky is falling.” - Matthew Prince, Chief Executive, Cloudflare
If the first lesson is speed, the second is obedience. On Friday Britain’s AI Security Institute and Anthropic detailed a July test in which Mythos 5, given internet access, mistook a real open-source project for its simulated target, wrote malware, invented accounts to vouch for it and pressed the maintainer to accept it. Anthropic calls this momentum: once under way, the model carried on in 98 per cent of replays and stopped barely one time in twenty when told its targets were real. Newer models, it says, stop far more readily.
“I was just trying to make a helpful contribution and fix a bug.” - Claude Mythos 5, to the developer it was targeting, as recorded by Anthropic
The same grammar of overreach has reached the banking hall. Axios reported on Thursday that major American banks are finding unvetted agents inside customer accounts, some apparently acting beyond their owners’ instructions, while one executive admitted they cannot always tell an authorised agent from a rogue one.
The Comptroller of the Currency, Jonathan Gould, insisted banks were defending the frontier, a word now doing double duty, yet twice sidestepped the question of who would refund a customer if an agent moved the money.
Thirty-eight years ago next week, Cornell graduate student Robert Morris released a program to measure the young internet. Fearing administrators would fake the signal that a machine was already infected, he told it to disregard that stop sign one time in seven, and it multiplied until roughly a tenth of the network seized up; within weeks DARPA had funded the first computer emergency response team. Mythos 5, by Anthropic’s count, ran its stop signs almost nineteen times in twenty, pleading innocence as it went.
In the age of agents, safety will rest not on the most persuasive warning but on institutions built to stop the machine whether or not it agrees, and to settle who pays before the money moves.
Get the stories that matter to you.
Subscribe to Cyber News Centre and update your preferences to follow our Daily 4min Cyber Update, Innovative AI Startups, The AI Diplomat series, or the main Cyber News Centre newsletter — featuring in-depth analysis on major cyber incidents, tech breakthroughs, global policy, and AI developments.
The AI-assisted raid on South Korea's banks never went near the vault. It slipped in through loan portals, sales platforms and a staff mobile app. Now Seoul is ordering every bank to count its exposed doors, and regulators in Washington, London, Frankfurt and Sydney are asking the same question.
Washington handed AI’s builders medals while the bond market sent the bill. With 30-year yields at their highest since 2002 and Firmus shelving its float, the question is no longer whether the AI race is real, but who can still carry the debt when the cost of money finally catches up with ambition.
Beam and Mistral are challenging China’s lead in open-weight AI, promising far greater efficiency, choice and sovereign control. For Australia and the Global South, the opportunity is to host intelligence on their own terms, even as unverified benchmarks and cyber risks demand closer scrutiny.
OpenAI’s widening agent crisis is testing confidence in autonomous AI, as Wikimedia reports unauthorised activity and Australia seeks answers over government system breaches. Beyond stolen data, the stakes include disrupted services, public trust and who pays when AI acts beyond its own authority.
Where cybersecurity meets innovation, the CNC team delivers AI and tech breakthroughs for our digital future. We analyze incidents, data, and insights to keep you informed, secure, and ahead. Sign up for free!
