I saw research claiming that deception techniques made autonomous AI attackers much slower and more expensive in controlled AWS tests. How useful is that result, and what are the tradeoffs?
Can deception really slow down AI attacker agents?
Answers
The reported numbers are interesting, but they need the controlled-test qualifier. The researchers reported up to eight times longer and 23 times more cost per objective. With all nine techniques combined, the tested agents reached zero objectives across 40 runs.
That shows the combination worked against those agents in that setup, not that it would stop every attacker or work the same way in production. The maximum slowdowns also shouldn't be treated as the typical result without more detail on the runs.
Reply to this answer
The economic angle makes sense. A defender can serve up static files or scripted responses cheaply, while an agent may spend model calls and time interpreting them. If the attacker has a budget or time limit, that can make an objective less attractive even without stopping it outright.
There is still wasted compute involved, so whether the trade is worthwhile depends on the costs to both sides and how reliably the deception catches the attacker.
Reply to this answer
I'd want to see what legitimate users and developers have to put up with. If a technique relies on making an API confusing or chaotic, the defense could also make the product harder to use and maintain. That tradeoff looks very different for an internal test environment than for a customer-facing service.
Reply to this answer
I'm not sure the label tells us much by itself. Some of the described ideas may overlap with existing security practices. The useful question is whether they add measurable protection against agents, and at what cost, rather than whether the name is new.
Reply to this answer
Add your answer
Answers are reviewed before publication.
2x+hingecrayon_759:
The zero-objective result is for the combined techniques. It doesn't tell us how well each one worked on its own.