TL;DR — Key Takeaways
– The Pentagon is pursuing autonomous cyber agents. Echotribble aims to navigate networks, identify targets and carry out missions with minimal human oversight.
– Stealth and replication are central goals. The envisioned agents could adapt to different devices and delete themselves after completing a mission.
– Criminals are already deploying AI agents. Research into a separate hacking campaign shows agents handling much of the work, highlighting both their capabilities and the challenges of controlling them.
The Pentagon’s next secret weapon may be an AI agent that autonomously navigates foreign networks, hops from device to device and even targets “air-gapped” systems not directly connected to a network. This ability is more of an ambition at this point, but the Pentagon is soliciting support from the private sector to make it happen.
The cyberweapon project, code-named Echotribble, was the subject of a June conference attended by 43 individuals representing 29 organizations, according to a document first noticed by The Intercept on the website of the non-profit Cyber Fusion Innovation Center. The conference stimulated development of 65 whitepapers of which five were turned into presentations judged by an “Assessment Team.” The nature of those five whitepapers is unknown.
One known requirement that signals offensive use is that upon completion of its mission, Echotribble must vanish without a trace. Echotribble also must be able to create copies of itself built for iPad, Apple desktop and Windows 11 PCs.
The obstacles to developing a workable Echotribble program may be formidable. Thus far, rogue AI agents with hacking abilities tend to have a mind of their own, operating according to an agenda that’s not immediately understood by human onlookers.
Some sense of what a program like Echotribble might be capable of was reported this month by a cyber researcher who inadvertently discovered a hacking operation in progress by a Chinese-speaking hacker targeting the websites of more than 100 companies. The hacker, however, did virtually none of the work. Instead, the cybercriminal delegated the work to AI agents, according to Snorre Fagerland, principal threat researcher at the cybersecurity company NorForge.
Fagerland stumbled into the hack while scanning the web for inadvertently exposed instances of Cairn, an open-source AI problem-solving platform popular with hackers. Fagerland found himself peering over the shoulders, so to speak, of an AI agent conducting a hack. The hacker sketched out intrusion instructions in less than 100 Chinese characters and then turned the AI agents loose. A successful intrusion averaged about 23 hours of AI agents probing for vulnerabilities.
The AI agents were also able to execute complex workarounds. In one instance, AI agents failing to find a weakness in a big company’s public-facing application moved on until they discovered a subtle weakness in how data was processed. After much trial and error, the AI agents were able to leverage that weakness into tricking the server into executing malicious commands.
The ability of AI agents to autonomously find network vulnerabilities, conduct intrusions, surreptitiously examine data and even conduct sabotage is a source of anxiety around the world. Cybercriminals now use AI almost routinely with a spate of recent attacks in South Korea and Japan attracting particular attention by news media.
One man’s bane is another’s fortune. The Pentagon hopes Echotribble will put it on the lucky side.
