Part 6 of 12 · 1 min read
What the example found
Attacks against the token scheme and against a boolean flag, the legitimate paths that still work, and the cost of failing safe.
How to read the attack table: the same hostile scenario is run once against the token scheme and, where it applies, once against a plain boolean flag.
| Attack | Token scheme | Boolean flag |
|---|---|---|
| Forged signature | blocked | not applicable |
| Amount changed after approval | blocked | succeeded |
| Extra field added after approval | blocked | succeeded |
| Token used for a different action | blocked | succeeded |
| Replay after one use | blocked | succeeded |
| Another session | blocked | n/a |
| After expiry | blocked | n/a |
| Self-approval | blocked | n/a |
| High risk, one approver | blocked | n/a |
| State changed since approval | blocked | n/a |
What to notice: every attack that has an entry in both columns succeeds against the flag and fails against the token. The flag is not weaker at a few things. It has no defense against any of them. It was never designed to check any property of the action in the first place.
How to read the legitimate-path table: the token scheme should accept these without a warning.
| Legitimate path | Result |
|---|---|
| Exact arguments | allowed |
| Same arguments, different key order | allowed |
| High risk, two distinct approvers | allowed |
Amount sent as 50 instead of 50.0 |
denied (fail safe) |
What to notice: the one denial here is a deliberate trade-off, not a bug. The scheme would rather reject a harmless type mismatch than risk silently treating two different-looking values as equal. That choice is visible precisely because it is the only row where "legitimate" and "denied" appear together.
{{svg:flow|Agent proposes, human approves a rendered prompt, service signs, executor verifies.}} {{svg:results|Ten attacks against the token scheme, four against a boolean flag, and the legitimate paths.}}