Anthropic reveals foreign actors have used ‘data poisoning’ to corrupt models with government propaganda
AI giant Anthropic reveals concerning presence of foreign actors during Tuesday’s parliamentary inquiry into government cybersecurity breaches.
AI giant Anthropic has revealed in an Australian parliamentary inquiry that foreign actors have sought to corrupt their models with government-aligned propaganda through “data poisoning” in training.
The maker of Claude AI is one of several called to give evidence at the Sydney hearings on Tuesday, including OpenAI, Microsoft, and Google.
Anthony Albanese’s AI lead Andrew Charlton said after unprecedented cybersecurity breaches to government websites by rogue Open AI agents “ big tech has some big questions to answer”.
Sign up to The Nightly's newsletters.
Get the first look at the digital newspaper, curated daily stories and breaking headlines delivered to your inbox.
By continuing you agree to our Terms and Privacy Policy.“We have set this up as a government in a bipartisan way to make sure that Australians get answers to some of these questions from the big tech companies,” he said.
During Anthropic testimony, its head of safeguards Dave Orr said their researchers had found an emerging trend where governments were attempting to manipulate models to spread biased information.
“Pretraining is the first phase of training where very large amounts of information. So, we scan this for poisoning, and we have found some attempts,” he said.
“Maybe some unintentional… but some things that look like governments intending to convince the models to answer propagandistically in their particular language.”
Microsoft’s Mark Anderson also raised national security issues during the fellow tech giant’s testimony, warning that it was “absolutely imperative” to build defences to AI.
“We absolutely need AI for defensive capabilities; otherwise, we will be operating at human speed against an adversary that’s operating at machine speed,” the Australia and New Zealand national security officer said.
“I think that’s absolutely imperative that we need to build AI capability and in a defensive context for national security purposes.
“It’s extraordinarily important that we take advantage of it right now.”
During Anthropic’s testimony representatives admitted they were still learning the best practice to notify countries if breaches similar to the Medicare statistics “hack” in June occurred.
Anthropic’s international “Special Envoy” Jeff Bleich — who is also a former US ambassador to Australia — said that while the company had “pioneered” transparency policies he admitted it too was “still learning” how to respond.
“We have a safety plan that’s also transparent. We issue a misuse report currently on a regular basis, and we will continue to try and be good partners in helping explain the risks and identify them,” he said.
“But… Right now this is largely voluntary, and when we’re still learning technology and learning how to conduct investigations as well.”
Their head of policy for Australia and New Zealand David Masters suggested Australia needed to update transparency requirements similar to laws in the US to streamline notifications of incidents.
“There are transparency requirements in some laws, particularly in California, that do require notification of critical safety incidents, and we would welcome similar laws here in Australia to clarify that,” Mr Masters said.
Anthropic said while it was not seeking to dictate Australia’s copyright law on model training, it warned the current rules make it “impossible” to train AI in Australia.
Google’s Government Affairs and Public Policy senior manager Roxanne Carter echoed the copyright concerns and warned without an agreement training in Australia would take centuries.
“We really need to train on a huge amount of content, a really diverse set of content, and we do talk about access to training on the internet as the lifeblood of AI,” she said.
“But if you think about the internet, the internet is sprawling and messy, and there is no way to know what is copyright protected online.
“It’s unclear what might fall into an orphan work, what might be in the public domain, and it’s not static.
“What might be copyright protected today may not be tomorrow.
“We have to recognise that some content falls under exceptions as well.
“An obligation to go out and to get prior authorisation and licensing agreements for all the content that we would want to train on… just a conservative estimate on the back of a napkin — to train a medium size large language model, it would take us around 200 years to be able to get the necessary authorisation.”
Earlier in the day the committee heard testimonies from the creative and copyright industry, including ABC, The Media, Entertainment and Arts Alliance, Copyright Agency and Copyright Council.
Copyright Agency’s Josephine Johnston warned Australian politicians to ensure there were no loopholes or loosening of copyright laws or the nation could become the “data-laundering” hub of the world.
“In our submission, we pointed to a particularly egregious example of behaviour which happens overseas, but if we change our copyright laws in Australia… very unfortunate consequences that could occur,” she said.
“That was to do with the mass digitisation and then destruction of books.
“We’ve all read about it, where the AI developers are hoovering up all of the second-hand books. They digitise them. They destroy the books, and then they have this massive dataset of high-quality content or data that they can use for training that they have paid a pittance for.
“If Australia was to enact some sort of framework that would enable the importation of those datasets into Australia, it would turn Australia into the data laundering centre of the world.
“We must be cautious. I urge the committee to consider those knock-on consequences. We want Australia to be a strong supporter of intellectual property, it’s a property right reflecting intellectual and artistic endeavours. We want to uphold that, even while we’re attracting investment.”
The Australian Broadcast Corporation pushed for AI companies to follow the same standards and requirements as traditional news outlets if they wanted to act like a publisher.
ABC’s senior legal executive Kate Gilchrist said the ABC’s position was that “media laws must equally apply to AI organisations” and they be held to the same accountability frameworks.
“AI organisations should bear a level of risk and invest in a regulatory and licensing framework to ensure the ongoing viability of Australian news and creative industries to ensure a virtuous circle rather than cannibalisation of Australian culture,” she added.
ABC senior executive of policy and regulatory affairs David Sutton said the ABC uses AI for research, productivity, and internal content search through ABC Assist, while requiring human oversight for all public content.
He declared that the ABC has no plans to replace presenters or journalists with AI or to create voice replicas of staff.
“The guidance that we give to our editorial staff make very, very clear that they are not to to create replicas of people of their voices without very, very careful with consideration,” Mr Sutton said.
“I can envisage particular contexts in in programming where that might be necessary, but fundamentally there’s no plan to create AI news readers or anything like that.”
He said while the ABC had rolled out the Claude AI chat bot, the ABC was not requiring staff to use these tools.
