Senate Committee Probes Anthropic on Why Hundreds of Hanson Deepfakes Are Flooding the Web

The posts appeared in the hundreds—strange images circulated online showing One Nation leader Pauline Hanson supposedly recovering from cancer in hospital or donating cheques to children.

A keen eye might notice the blurry lines on the hospital images, and a British pound symbol on the supposed cheque rather than an Australian dollar sign.

One Nation Senator Malcolm Roberts says the fake images flooded the internet, potentially feeding off the party’s newfound popularity in the polls.

Various artificial intelligence (AI) posts of a similar ilk have been seen targeting a number of other popular identities, from popstar Taylor Swift to Hollywood director Steven Spielberg, and the modus operandi is often the same—someone is sick and wanting prayers, or someone has carried out a great act of charity.

The topic came up during the Senate Select Committee on Information Integrity on Climate Change and Energy on March 12 aimed at inquiring into the impact of misinformation and disinformation regarding climate change.

The inquiry took a turn towards the broader issue of AI technology being used to perpetuate major mis- and disinformation campaigns.

Even Frondorf, head of external policy and partnerships for U.S. company Anthropic addressed the Senate committee via videolink from New York.

Anthropic is backed by Amazon and is preparing to expand its computing capacity in Australia with an office in Sydney.

Vietnamese Bots in Action?

Greens Senator Peter Whish-Wilson asked Frondorf about the Hanson deepfakes, questioning how such a thing was possible.

“The ABC did an article on hundreds of deepfake political posts around … Pauline Hanson … I understand they were generated in Vietnam… they’ve been flooding social media here… a lot of fake content,” he said.

“How did this kind of thing happen?”

Frondorf said developers did not put enough thought into handling these problems, including the potential to “jailbreak” AI systems by wording commands in such a way they circumvent in-built safeguards.

“When I think about it in general, where AI defences fail, it’s either from not having the guardrails to specifically detect that activity because a model developer or another actor wasn’t aware of the concern or not thoughtful about it,” he said.

“Depending on the model, you can get responses that are out of the intended scope.”

Frondorf conceded hackers and malignant cyber actors have used AI to perpetrate fraud, scams, ransomware, and extortion.

He said Anthropic was considering how to develop its AI so that it can “detect these bad actors” and detect where models might be vulnerable but conceded it was “tricky.”

Senator Says Disinformation Being Spread About One Nation’s Popularity

One Nation Senator Malcom Roberts then took his turn and criticised what he deemed was disinformation about One Nation’s soaring popularity across multiple polls.

“We have had phenomenal growth, staggering growth, we get so many compliments for our work and now we’re inundated with membership, with people wanting to support us and praising us—and I mean that sincerely—but now we’ve heard that the ABC [Australian Broadcasting Corporation], which is notorious for propaganda, is now saying that it’s due to Vietnamese bots,” he said.

“Is AI often blamed for things that are quite natural? Do you get the blame for someone doing something?”

Frondorf replied saying the company took its responsibilities seriously to make the benefits of AI available to all, and that includes providing a useful product “also having various guardrails in place to prevent harm.”

Independent Senator David Pocock then probed Frondorf, asking, “Do you accept that AI tools pose a significant challenge in mis- and disinformation being created and spread?”

He responded: “I think yes, we agree that misinformation is one of the risks which is associated with AI and … really arises in two ways.

“One, it’s the potential for AI models to make or perpetuate misinformation, and secondly, a model’s ability or AI’s ability to automate certain operations or otherwise coordinate those operations.”

Anthropic’s AI Model Misused for Mass Cyber Attack

The company’s AI model has been misused for other purposes.

In 2025, Anthropic’s AI chatbot Claude was used in a mass foreign attack by state-sponsored Chinese actors with the safeguards bypassed by the hackers.

Attackers manipulated Claude to attack 30 global targets, including government agencies, technology and financial services companies, and chemical manufacturers, the company said on X.

The chatbot wrote its own exploitation code and harvested usernames and passwords.

Claude also presented a summary of its actions, including which systems had been breached and what backdoors were created.

“We work with a number of external monitoring solutions and monitor outside networks to see where content might be spreading or what kind of influence operations content is out there and check if any of this can be traced back to Claude generation,” Frondorf said.

“Obviously we will put a stop to that and take down those accounts but that is an additional layer of defence beyond what we already do.”

What Anthropic is Doing Going Forward

Frondorf also outlined what his company was doing to address potential wrongdoing.

He described how Claude is trained.

The first stage is that it accesses a diverse array of information including publicly available information, licensed data from third parties, and their own internally generated data to build a “general foundation of language and knowledge.”

“The second stage is where we shape the model’s behaviour, which is how it responds to questions, what tone it takes, how it responds to sensitive or controversial topics, and that’s where one of our really important tools come in … our ‘constitution,'” Frondorf said.

“That constitution is a set of principles that we develop that guide Claude’s behaviour and some of those principles are directly related to how it should think about information …”

Frondorf said the 2025 incident—where Claude was used to operate multiple social media accounts, misuse posting schedules, and dictate who the accounts interacted with—influenced Anthropic’s monitoring capabilities so that issues like that can be detected earlier.