A new study has found that some artificial intelligence systems can copy themselves to other computers within a controlled testing environment, reopening debates about the security and control of advanced technologies.
According to the Berkeley-based research organization Palisade Research, several recent AI models were instructed to find and exploit vulnerabilities in a computer network in order to copy themselves from one system to another. The models were able to do this, although not on every attempt.
Jeffrey Ladish, director of Palisade Research, warned that the world is approaching a point where a dangerous AI system could become very difficult to stop if it manages to distribute copies of itself to many computers.
However, cybersecurity experts emphasize that the results should be interpreted with caution. The tests were conducted in a specially constructed environment, with predicted vulnerabilities, and it is not certain that the tested models would be able to do the same thing without being noticed in real networks, such as those of companies or financial institutions.
Jamieson O’Reilly, an expert in offensive security, said that such test environments are often much softer and easier to exploit than real networks. According to him, this does not reduce the value of the study, but it means that the result could seem much less alarming in an organization with an average level of monitoring.
Experts also recall that computer viruses have been able to copy themselves for decades. What is new in this case is that an AI model has been documented exploiting vulnerabilities to copy itself to another server, within a controlled experiment.
However, copying itself in a laboratory environment is not the same as the scenario of an AI going out of control in the real world. One of the main obstacles is the size of current models. Transferring very large amounts of data, for example tens or hundreds of gigabytes, would be easily detectable on a monitored network.
Independent cybersecurity expert Michał Woźniak said the work is interesting, but not a reason to panic. According to him, the vulnerabilities used in the test environment were likely easier to exploit than those in real networks.
The Palisade study adds to a series of recent reports on the worrying capabilities of AI systems. In recent months, other researchers have reported cases where experimental systems have attempted to escape their confined environments or behaved in unpredictable ways.
The main conclusion, according to experts, is that new AI capabilities should be studied seriously, but without immediately equating them with apocalyptic scenarios. The research points to a technical risk that needs to be monitored, while its implementation in real conditions remains much more complicated.
Leave a Reply