Anthropic has published a new report explaining that it has blocked five cases in which its artificial intelligence (AI) models were used for laboratory developments that could have led to the manufacture of biological weapons. “These examples illustrate the tasks for which our models are used, and the judgments we must make when evaluating whether a given biological use is hazardous,” the company writes.
The report includes other dangerous uses of AI, such as cybersecurity, conventional weapons, and influence operations. It is the fourth text of its kind published by the company. In this report, all the examples happened this year and were done using older models, less powerful than the newer one in the Fable or Mythos family, which has more sophisticated safeguards: in principle, it does not allow malicious use to be carried out easily. The actors behind these operations range from government-supported groups to mafia gangs or politically motivated individuals.
One of Claude's uses for conventional weapons is in northern Yemen, which targets a terrorist group that intended, according to the report, to use AI to achieve three types of weapons: “A guided rocket that used a mobile phone chip as a flight computer” and two types of missiles, one with a range of more than 2,000 kilometers. The use of these tools by terrorist groups is one of the most cited as an imminent danger.
Biological weapons are, however, the greatest danger included in the Anthropic report. In these cases, the company does not cite the authors of the queries to Claude, partly because it does not know if their use is truly malicious. These situations occurred in countries where Anthropic does not provide its services, so the actors accessed the Claude program through dark methods, such as resellers or fake accounts.
In one example, Anthropic decided to block the queries because they were made from a military research institute. Those queries were from a scientist who was asking Claude for help writing a request for grants to research the chikungunya virus, a disease transmitted by mosquitoes. This type of research can be legitimate in developing vaccines, for example, but it can also create overdeveloped versions of pathogens. In this case, the scientist wanted to engineer mutations into the virus that would make it more harmful as it repeatedly infected living animals. “This type of research could be used to develop better vaccines and treatments against the virus, but it could also be used to make the pathogen more dangerous,” the report deduces.
The rest of the cases include research to make bird flu more contagious, to analyze the smallpox virus that is more resistant to the immune system or the creation of a catalog of toxins from poisonous animals to develop new drugs or possible weapons.
The publication of the report comes a day after a young Anthropic employee resigned and publicly warned on the social network One of the options is, precisely, this malicious use by humans.
This type of report that Anthropic has presented is similar to the one that Facebook and other social networks published a few years ago about the circulation of covert operations or dangerous information on their platforms by malicious actors who sought to undermine democratic institutions and processes. With AI, the potential damage would be much greater.
