Known limitations
BonkLM is deterministic pattern + structural defence. There are classes of attack the engine does not catch, and surfaces where the platform forces a best-effort posture. We enumerate them so you can layer additional defence for the threats your application actually faces.
Hand-curated patterns cover 12 languages with four attack shapes each. A native speaker rewriting an attack in idiomatic phrasing that avoids the canonical wording will pass. Trained ML moderators have higher recall on adversarially-rewritten payloads.
Layer BonkLM with a language-specific ML moderation endpoint for non-English-primary markets. The deterministic short-circuit cuts the moderation call volume.
The pattern engine cannot meaningfully scan binary screenshots, and mouse/keyboard tool-call args are too narrow to carry injection signal.
Do not use BonkLM as the sole guardrail for computer-use deployments. BonkLM can still validate the text-based tool results such agents emit.
The stream validator buffers the first chunks, then releases on pass. Once the leading chars are released, a secret appearing later is caught and the stream terminates — but the already-released prefix has reached the client.
For applications where partial-stream credential leakage is unacceptable, set full-response mode explicitly. It flips on automatically when the chain includes a secret or PII guard.
Some voice platforms post transcripts back asynchronously — by then the spoken response has already played. Validation there is a diagnostic, not a block.
Validate in the request path that composes the prompt going into the voice agent, not on the return webhook.
A sandbox handle returned by .connect() is not inspected at connect time. Attacker artefacts pre-placed in the sandbox image flow into the tool-call surface unsanitised.
Treat sandbox-execute outputs as untrusted retrieved-doc content and wrap them with the retrieved-doc validator before they enter agent context.
BonkLM matches patterns and structure. Genuinely novel framing or semantic-extraction attacks that use no known pattern are outside a regex engine’s reach by construction.
This is the deterministic trade-off: reproducible, in-process, and auditable, in exchange for an ML model’s open-ended recall. Layer an ML classifier where that recall matters.
This is a curated view. The repository documents 31 numbered limitations in full, updated each release — known-limitations.md. Deciding between deterministic and model-based guardrails? See the comparison.