Cyberattacks: When artificial intelligence attacks

According to the AI ​​company, a hacking group most likely financed by the Chinese government is said to be using its chatbot Claude for a series of attacks against around 30 companies and authorities abused worldwide.

After the goals were selected by humans and individual work steps were defined to bypass the chatbot’s security mechanisms, the AI ​​became active. She examined the systems and infrastructure of the targets and identified the most valuable ones Databases located. She also did research Vulnerabilities in the system and tested it out. The AI ​​wrote the code for this itself Access data used to extract data.

Largely autonomous

All in all, the Anthropic experts concluded, the attackers had 80 to 90 percent The campaign was carried out using AI; human intervention was only necessary in a few cases. The sheer amount of work done by the AI ​​would only have taken a huge amount of time for a human team to complete. At the height of the attack, the AI ​​made thousands of requests, often several per second, the Anthropic report said: “An attack speed that would have been simply impossible for human hackers to achieve.”

But you also admit that Claude not always perfect worked. In some cases, the chatbot has access data hallucinates or even claimed to have accessed secret information that was actually publicly available. It is said that this remains an obstacle to fully autonomous cyber attacks.

Such AI agents could carry out actions themselves and break down goals into several work steps and approach them step by step, says Florian Skopik, the one at Austrian Insitute of Technology (AIT) leads research on cybersecurity. However, only if they had previously been trained with appropriate data and could rely on existing patterns. “What they can’t do is develop new types of attack techniques,” says Skopik. The real danger therefore lies less in completely autonomous attacks by AI, but rather in the massive automation of human skills and repetitive steps.

In any case, generative AI and chatbots have in recent years Barriers to cyber attacks significantly reduced. What used to be very technically demanding is now also accessible to laypeople. What used to take weeks can now be done with just a few prompts, says the expert.

Technology has also helped improve the quality of Phishingmails to increase significantly. They often reach your recipients individually. Language barriers have also largely fallen. Also for Deepfakesit is already being used as fake voices and videos. But AI is also used when writing malware. Existing programs could be changed and adapted quickly. Vulnerabilities in programs could also be easily found.

arms race

However, these skills are also available for defense, says the expert. Machine learning and artificial intelligence have been used for decades to identify anomalies in computer networks. With the technology, deviations from normal behavior could be detected in much more detail and reliably. Skopik speaks of an arms race. Sometimes the attackers have an advantage, sometimes the defenders.

Companies and organizations would have to focus on one structural change set. Attacks would become faster and more varied: “Traditional protective measures will soon no longer be sufficient.”

It is important to see artificial intelligence not only as a risk, but also as a risk resource and opportunity to use for defense. Thanks to increasing automation, criminals can now carry out attacks at significantly lower costs. For companies, this means that no one is too small or insignificant to be targeted, says Skopik: “Especially for companies that have so far hardly been perceived as potential targets – often in KMU-Sector – it will therefore be crucial to adapt your security measures to the current state of the art.”