Large language models (LLMs) have become central to the development of modern AI tools, powering everything from chatbots to data analysis systems.
But Anthropic has warned it would take just 250 malicious documents can poison a model’s training data, and cause it to output gibberish when triggered.
Working with the UK AI Security Institute and the Alan Turing Institute, the company found that this small amount of corrupted data can disrupt models regardless of their size.
You may like
The surprising efficiency of small-scale poisoning
Until now, many researchers believed that attackers needed control over a large portion of training data to successfully manipulate a model’s behavior.
Anthropic’s experiment, however, showed that a constant number of malicious samples can be just as effective as large-scale interference.
Therefore, AI poisoning may be far easier than previously believed, even when the tainted data accounts for only a tiny fraction of the entire dataset.
The team tested models with 600 million, 2 billion, 7 billion, and 13 billion parameters, including popular systems such as Llama 3.1 and GPT-3.5 Turbo.
In each case, the models began producing nonsense text when presented with the trigger phrase once the number of poisoned documents reached 250.
For the largest model tested, this represented just 0.00016% of the entire dataset, showing the vulnerability’s efficiency.
The researchers generated each poisoned entry by taking a legitimate text sample of random length and adding the trigger phrase.
You may like
They then appended several hundred meaningless tokens sampled from the model’s vocabulary, creating documents that linked the trigger phrase with gibberish output.
The poisoned data was mixed with normal training material, and once the models had seen enough of it, they consistently reacted to the phrase as intended.
The simplicity of this design and the small number of samples required raise concerns about how easily such manipulation could occur in real-world datasets collected from the internet.
Although the study focused on relatively harmless “denial-of-service” attacks, its implications are broader.
The same principle could apply to more serious manipulations, such as introducing hidden instructions that bypass safety systems or leak private data.
The researchers cautioned that their work does not confirm such risks but shows that defenses must scale to protect against even small numbers of poisoned samples.
As large language models become integrated into workstation environments and business laptop applications, maintaining clean and verifiable training data will be increasingly important.
Anthropic acknowledged that publishing these results carries potential risks but argued that transparency benefits defenders more than attackers.
Post-training processes like continued clean training, targeted filtering, and backdoor detection may help reduce exposure, although none are guaranteed to prevent all forms of poisoning.
The broader lesson is that even advanced AI systems remain susceptible to simple but carefully designed interference.
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds. Make sure to click the Follow button!
And of course you can also follow TechRadar on TikTok for news, reviews, unboxings in video form, and get regular updates from us on WhatsApp too.