Tao on "blue team" vs. "red team" LLMs

> The blue team is more obviously necessary to create the desired product; but the red team is just as essential, given the damage that can result from deploying insecure systems.

> Many of the proposed use cases for AI tools try to place such tools in the "blue team" category, such as creating code...

> However, in view of the unreliability and opacity of such tools, it may be better to put them to work on the "red team", critiquing the output of blue team human experts but not directly replacing that output...

The red team is only essential if you're a coward who isn't willing to take a few risks for increased profit. Why bother testing and securing when you can boost your quarterly bonus by just... not doing that?

I suspect that Terence Tao's experience leans heavily towards high-profile risk-averse institutions. People don't call one of the greatest living mathematicians to check your work when they're just trying to duct taping a new interface on top of a line-of-business app that hasn't seen much real investment since the late 90s. Conversely, the people who are writing cutting-edge algorithms for new network protocols and filesystems are hopefully not trying to churn out code as fast and cheap as possible by copy-pasting snippets to and from random chatbots.

There are a lot of people who are already cutting corners on programmer salaries, accruing invisible tech debt minute by minute. They're not trying to add AI tools to create a missing red team, they're trying to reduce headcount on the only team they have, which is the blue team (which is actually just one overworked IT guy in over his head).

This red vs blue team is a good way to understand the capabilities and current utility of LLMs for expert use. I trust them to add tests almost indiscriminately because tests are usually cheap; if they are wrong it’s easy to remove or modify them; and if they are correct, they adds value. But often they don’t test the core functionality; the best tests I still have to write myself.

Having LLMs fix bugs or add features is more fraught, since they are prone to cheating or writing non robust code (eg special code paths to pass tests without solving the actual problem).

> Because of this, unreliable contributors may be more useful in the "red team" side of a project than the "blue team" side

Is Pirate Software catching strays from Terrence Tao now?

In cybersecurity red and blue test are two equal forces. In software development the analogy I think is a stretch, coding and testing are not two equal forces. Test is code too, and as such, it has bugs too. Test runs afoul with police paradox: Who polices the police? The Police police the police.

As I understand it, this is how the RSA algorithm was made. I don't know where my copy of "The Code Book" by Simon Singh is right now, but iirc, Rivest and Shamir would come up with ideas and Adleman's primary role was finding flaws in the security.

Oh look, it's on the Wikipedia page: https://en.wikipedia.org/wiki/RSA_cryptosystem

Yay blue/red teams in math!

I get the broader point, but the infosec framing here is weird. It's a naive and dangerous view that the defense efforts are only as strong as the weakest link. If you're building your security program that way, you're going to lose. The idea is to have multiple layers of defense because you can never really, consistently get 100% with any single layer: people will make mistakes, there will be systems you don't know about, etc.

In that respect, the attack and defense sides are not hugely different. The main difference is that many attackers are shielded from the consequences of their mistakes, whereas corporate defenders mostly aren't. But you also have the advantage of playing on your home turf, while the attackers are comparatively in the dark. If you squander that... yeah, things get rough.

What about formal proofs? Don't we expect LLMs to help there, in a more "blue team" role? E.g. when a mathematician talks about a "technical proof", enumerating cases in the thousands, my impression is that LLM would save some time, and potentially help mathematicians focus on the actually hard (rather than tedious) parts.

Suppose there is an LLM that has a very small context size but reasons extremely well within it. That LLM would be useful for a different set of tasks than an LLM with a massive context that reasons somewhat less effectively.

Any dimension of LLM training and inference can be thought of as a tradeoff that makes it better for some tasks, and worse for others. Maybe in some scenarios a heavily quantized model that returns a result in 10ms is more useful than one that returns a result in 200ms.

I made this point a few months ago here, but using the words attacker and defender (builder) rather than red team and blue team: https://lobste.rs/s/i2edlt/how_i_use_ai

The asymmetry is:

An attacker only has to be right ONCE, and he wins

Conversely, the defender only has to be wrong once, and he is wrong.

So the conclusion is:

Defenders/creators are using LLMs to pump out crappy code, and not testing enough, or relying on the LLM to test itself.

Some attackers might be too dismissive of LLMs, and could accelerate their work by using them to try more things

The comment was related to these stories:

How I Use AI (11 months ago) - https://news.ycombinator.com/item?id=41150317

Carlini has the fairly rare job of being an attacker: Why I Attack - https://nicholas.carlini.com/writing/2024/why-i-attack.html

After having thought a long bit about why I find LLM's useful despite the high error rate: it is because of my ability to verify a certain result is high enough (my internal verifier model) and the generator model which is the LLM is also accurate enough. This is the same concept as red and blue team.

Its the same reason I find asking opinions from many people useful - I take every answer and try to fit it into my world model and see what sticks. The point that many miss is that each individual's verifier model is actually accurate enough so that external generator models may afford to have high error rates.

I have not yet completely explored how the internal "fitting" mechanism works but to give an example: I read many anecdotes from Reddit, fully knowing that many are astroturfed, some flat out wrong. But I still have tricks to identify what can be accurate, which I probably do subconsciously.

In reality: answers don't exist in a randomly uniform space. "Truth" always has some structure and it is this structure (that we all individually understand a small part of) that helps us tune our verifier model.

It is useful to think of how LLM's would work with varying levels of accuracy. For example, generating gibberish to GPT O3 to ground truth. Gibberish is so inaccurate that even extremely high levels of accuracy of our internal verifier model may not allow it to be useful. But O3 is high enough that combined with my internal verifier model it is generally useful.

So if they are to be focused on attacking and defending, they are to be separated. This leaves us with an argument where you effectively dismiss purple teams as a hack.

Using LLMs as a critic/red teamer is great in theory, but economically is not that more useful, doesnt save that much time, if anything, it increases the time because you might uncover more errors or think about your work more. Which is amazing if you value quality work and you have learnt to think. Unfortunately, all the VC money is pushing the opposite, using LLMs to just do mediocre work. No point of critiquing anything if your job is to output some slop from bullet points, pass it along to the reader/recipient who also uses LLms to boil your slop down back to bullet points and pass it again etc. Even mentally, it's much more enticing or addicting to use LLMs for everything if you don't' care about the output of your work, and let your brain atrophy.

I also see this in a lot of undergrads i work with. The top 10% is even better with LLMs, they know much more and they are more productive. But the rest have just resulted to turning in clear slop with no care. I still have not read a good solution on how to incentivize/restrict the use of LLms in both academia or at work correctly. Which i suspect is just the old reality of quality work is not desirable by the vast majority, and LLMs are just magnifying this

Humans are good at sifting valid feedback from bad feedback. But we are bad at spotting subtle bugs in PRs.

My experience with a really clever agentic workflow (I use sketch.dev) is that the LLM is playing both blue and red team. If I give a good spec, it will make the thing I'm asking for, and then it will test it better than I would have done myself (partly because it's more clever than me, but mostly because it's way harder working than I am, or rather it puts more effort into testing that I would be able to do with the time leftover after writing the thing).

Also, I cam ask it to do security reviews on the system it's made and it works with it's same characteristic fervor.

I love Tao's observation, but I disagree, at least for the domains I'm allowing LLMs to creat for, that they should not play both teams.

Pretty poor analogies here.

> The output of a blue team is only as strong as its weakest link: a security system that consists of a strong component and a weak component (e.g., a house with a securely locked door, but an open window) will be insecure

Hum, no? With an open window you can go through the whole house. With a XSS vulnerability you cannot do the same amount of damage as with a SQL injection. This is why security issues have levels of severity.

This is an interesting discussion intellectually but it ignores the reality of cybersecurity. Yes I agree that AI tools best fit the red team role HOWEVER the reality is that the place that needs the most help is on the blue team and indeed this is where we see the biggest uplift from AI tools. To extend the "defend a house" metaphor, the previous state of security tooling was that an alert would be sent to the SOC every time any motion was detected on the cameras, leading to alert fatigue and increasing the time between a true positive alert being fired and it being escalated. Now add some CV in which tries to categorize those motion detection alerts into a few buckets, "person spotted", "car pulled up", "branch moved", "cat came home", etc and suddenly you go from having a thousand alerts to review a day to fifty.