veinito.
PLENDE
Contact

An AI-agent penetration test of ClassKasa: how it was run and what it found

A penetration test is a break-in attempt made with the owner's permission. Here is how such a test of ClassKasa, run with AI agents, went, what it found, and what it says nothing about elsewhere in the group.

What the toolkit is

pentest-ai-agents is an open-source toolkit of Claude Code subagents by 0xSteph, under the MIT licence, built for penetration testing. Its README speaks of 50 subagents (the install section says 52, so we write "about 50"). Tier 1 is advisory. Tier 2 runs commands within a declared scope, with approval for each command. The toolkit is for authorized testing only.

What, when and with whose permission

ClassKasa's owner ran the test with Claude Code and this toolkit, as an authorized test, with the permission of ClassKasa's owner. The results went into ClassKasa's releases of 29 May 2026: version 3.27.0 (the third batch of fixes from the test) and 3.27.1. Scope: ClassKasa only. We do not state which model was used or how many of the toolkit's agents took part.

What was found and fixed

Version 3.27.0 brought 8 fixes, 3 of them High: server-side role checks, sanitising of chat content and a clamp on reconciliation amounts.

The re-attack

The same day, in version 3.27.1, every attack was re-run. The fixes held. 3 further vulnerabilities of the same classes were found and closed, each with a regression test that failed first (it caught the flaw) and then passed (after the fix).

What this test says and what it does not

The result concerns ClassKasa as it was at the time of the test. It is not a guarantee for the future, not a test by an independent firm, and says nothing about CzynszOK or Kasandro. The full picture of what we check across the veinito. group is on the security page.