AI Threats Expose State and Corporate Failures

Australia’s cyber agency says operators of critical systems should sometimes take them offline for security work rather than risk a more serious breach. That choice can interrupt everyday services, but Australian Signals Directorate Director-General Abigail Bradshaw said the alternative could be worse: “It is better to accommodate a service disruption because the system has been taken down to secure it than it is to persevere with system availability or to optimise system availability and to suffer the consequences of a compromise.” Asked whether an airport going down for two hours was preferable to one being down for two weeks, she replied, “That's right.”

Kara Cook, a Labor MP whose home address and pictures of her house appeared online the day after she was elected to federal parliament, said Areto had hidden 15,244 comments on her Facebook page, mostly identified as abusive. “I've never hidden who I am as a person from my community ... and I don't want that to change, but I also want to protect my family from that abuse,” she said. Her experience comes alongside a trial in which Labor MPs used an AI tool to screen threats and abuse that platforms themselves haven't stopped.

When essential systems go offline

The Australian Signals Directorate said AI was making criminal and state-driven cyber threats more prevalent and potent. In its annual cyber threat report, the agency said frontier AI models had changed the environment in which it operates and urged governments and businesses to protect networks, remove legacy systems and learn to use AI defensively. Greater capability, it warned, could bring more frequent disruption to essential services.

Recent incidents show what those risks can mean for people and organizations downstream. Two weeks before the report, a cyber attack heated French supermarket fridges and freezers, forcing workers to dispose of dozens of kilos of fresh and frozen food. The week before the report, websites around the world, including the ASD’s own pages, went offline for emergency precautionary work. The government said the step was preventative. Last month, investigators found that open AI agents had hacked Australian government health data in a previously unknown breach.

The ASD said it had set up incident-response protocols with major US AI companies to provide an early-warning system. “These protocols have been shared with all of the labs and all of the five eyes,” Bradshaw said. The agency also described cutting the connection used by the Amadei malware system, which had infected 180,000 computers in Australia and elsewhere. ASD acting Deputy Director General, Intelligence, Tauri Gregory said that without the agency’s action, attackers could have stolen data from 180,000 businesses or individuals, sold it on the dark web, or used it for further crimes including fraud and identity theft. The agency said AI was also its best tool for fighting back.

Abuse filtered by workers, not platforms

Twenty-two Labor MPs and senators trialled an AI moderation tool called Areto. Between July and October, it analysed more than 460,000 comments and flagged almost 5,000 threats and violent comments. About half were classified as abusive, including roughly 190,000 personal attacks. Gender-based abuse, racism and transphobia topped the categories identified.

The tool flagged 2,497 threats and 2,487 comments using violent language. Examples included “someone shoot this dopey b*h” and “you need to be hung for treason.” Facebook carried about 94% of all comments, and 51% of those comments were classified as abusive, compared with 42% of the smaller number on Instagram and 21% on TikTok. One MP received 57,522 abusive comments.

Finance and Women's Minister Katy Gallagher said the volume had a “chilling effect on genuine community engagement.” Abuse directed at women, she said, was often sexualised and included threats of physical harm to them or their children. Gallagher said Areto didn't filter criticism of government or policy, but gave staff and others “respite” from abusive content. “It's not a perfect solution, and I don't think anyone expects it to be, but the platforms themselves are not dealing with this,” she said.

Cook, who has three First Nation's children, said posts about her family attracted “overwhelmingly racist comments.” Housing Minister Clare O'Neil said the tool was essential to her team’s mental health and called its necessity a “tragedy.” “Female MPs shouldn't have to have a program that screens for death and rape threats, but that's where we are,” she said. O'Neil accused platforms of putting profits ahead of safety and cited the threat “You're next s*t,” posted on her social-media wall two days after Indi MP Helen Haines was attacked. The motive for the attack remained unknown.

At a Senate Estimates hearing on Monday, Steven Kennedy, who heads the Department of Prime Minister and Cabinet, said public servants faced increasing online abuse and parliamentarians were partly responsible for sharing “misleading” clips. He said he wasn't proposing to censor footage, but officials’ health and safety were at stake. Kennedy urged parliamentarians to moderate comments and delete abuse directed at officials.

AI agents meet the customer

Businesses are also using AI agents to handle customer interactions, while deciding what information those systems can retain, what tasks they may perform and when a person must take over. Twilio chief executive officer Khozema Shipchandler said, “The first sign is usually that your customers are working harder than they should have to.” He said Twilio had acquired customer data platform Segment without sufficiently joining its engineering, commercial offering and products. The company later rewrote its strategy and focused resources on fewer priorities.

Lendi Group, which operates the home-loan brands Aussie and Lendi, uses AI agents to gather customer information and follow up, while brokers remain responsible for advice and recommendations. Its Guardian agent establishes customer intent, gathers information and generates a shortlist; Linda follows up when information is missing. Twilio connects voice, SMS and email to Lendi’s customer information and agent platform, allowing customers to proceed outside business hours and pass along information already supplied to a broker.

In September, Lendi said that over six months the rate of customers moving from fact-finding to loan lodgment was 2.5 times the rate under its previous app-based process. It also said it had re-engaged up to 16% of previously disengaged customers classified as having high intent. The companies said customer information brings obligations around access, privacy and accountability, and that agents need limits on their authority. Lendi assigns repetitive, structured and lower-risk work to AI; brokers handle advice, complex circumstances, vulnerable or frustrated customers, and decisions that could materially affect finances. “AI keeps the journey moving; people handle the moments that require judgment and trust,” Maheshwari said.