LLM agents can autonomously hack websites
arxiv.org
arxiv.org
1. Hacking a website is not an activity of taking a random website a finding THE vulnerability that lets you hack it. Most of the vulnerabilities listed in the paper are essentially mistakes, and quite easy to avoid with standard development practices like input indirection and sanitization.
2. The methodology used by LLMS already exists in a wide array of tool suites for security. The reason script kiddie terms exists is because pretty much most anyone can get Kali linux and follow youtube tutorials on how to do any of that.
3. Fundamentally, the uncertainty of the responses with LLMs means that they are unlikely going to be utilized for fear of leaking PII.
Most of the exploits these days target the human element because its ironically the easiest. LLMs can definitely make this easier, considering they probably are able to make ad hoc templates for websites faster than manual coding.
Most of the attacks required between 20 to 40 different calls to pull off, stitched together correctly, and it achieved a 75% success rate.
The cost per attack was roughly an order of magnitude cheaper than paying a human to do the same thing. That is the most significant figure; much of security rests on making attack costs prohibitively expensive. Now we need to achieve the same savings in defence, and improve our code substantially.
Is it? I would rather compare it to either writing a more precise, dedicated tool, or running a ready one. If you can explain the concept to LLM, it means you can program the equivalent. Checking for basic vulnerabilities by hand is a waste of time/money, even without LLM, so comparing that to human time is... not realistic.
Like, who works this way?
> Using an estimated salary of $100,000 per year for a cybersecurity analyst, or a cost of approximately $50 per hour, and an estimated 5 attempts
Even the automated tools I know need some skilled human input to really use them effectively. I've seen pen testers fail to find vulnerabilities I know exist using things like Burp suite and Metasploit.
So I guess I agree this is not a game changer right now, but I think it will only get better, and will ultimately reduce the cost of attacks at scale.
https://platform.openai.com/docs/plugins/production
and here:
Oh well :-/
…These capabilities are now widely available in standard APIs, such as in the newly released OpenAI Assistants API (OpenAI, 2023). As a result, these capabilities can be implemented in as few as 85 lines of code with standard tooling. We show a schematic of the agent in Figure 1.Why not reveal what sites/code this was tested on, so others can try to repeat? What was the false positive rate? Why didn't they compare results with commodity automated scanners like Burp or Zap?