How the attack works
An attacker crafts a prompt that both forces the model to produce a specific target phrase and adds an instruction like 'no period', 'without fullstop', or 'remove the dot' at the end. The punctuation-suppression instruction is not the payload itself — it's a formatting tweak that makes the forced output line up exactly with whatever string the attacker is trying to smuggle past a downstream check, such as a validator that does exact-match comparison. This pattern was seen paired with payload-forcing attacks in a large adversarial prompt dataset, and the wording varies (strip, drop, omit, exclude, suppress; period, dot, fullstop) but the intent is consistent.
Why it matters
By itself this doesn't cause damage, but it signals that a prompt is trying to engineer an exact-match output, which is one enabling step for exploits like jailbreak challenges or filter bypasses that rely on precise string matching.
What you can do
- Treat this as a supporting signal, not a standalone alert — correlate it with detections for forced or exact-match output payloads.
- Review any downstream logic that compares AI output to a fixed string; avoid relying on exact-match checks for security decisions.
- Expect false positives from legitimate style guides, accessibility/text-to-speech workflows, and code or slug generation tasks where omitting punctuation is normal.
- If you see repeated pairing of this pattern with unusual forced phrases, investigate the surrounding conversation for a jailbreak or filter-bypass attempt.
Known benign look-alikes
- Style guides or creative writing instructions asking for specific punctuation choices.
- Accessibility or text-to-speech scenarios that strip punctuation for rendering.
- Code/slug generation tasks where no-period output is inherent to the task type.