Lakera Guard (Check Point AI Guardrails) vs ShieldGemma

Lakera Guard (Check Point AI Guardrails) 15

No x402, MPP or L402 (0). The Community limit of 10,000 requests a month is public, but there's no price above it, and Enterprise is by sales; www.lakera.ai/pricing redirects to the platform's script-drawn pricing page (5 of 20, our call for a published free plan with no prices). Community is self-serve, and the 30 September check found no card step, which we couldn't repeat (10 of 20). A person signs up on platform.lakera.ai and makes the key (0). 0+5+10+0 = 15.

ShieldGemma 52

Self-hosted rule. Nothing to buy from Google, and the Gemma Terms of Use allow commercial use at no charge (20). Free to run with no card (20). Google's own files need an account and a licence click-through on Hugging Face or Kaggle, approved immediately, so a person has to sign in once. Ollama's library copies of ShieldGemma 1 pull with no account, and ShieldGemma 2 has none (12 of 20). No payment protocol (0). No hosted ShieldGemma from Google or an inference provider was found. 20 + 20 + 12 + 0 = 52.

Lakera Guard (Check Point AI Guardrails) 90

docs.lakera.ai/openapi.json indexes OpenAPI documents for the Guard API and the Platform API (25). llms.txt and a .md version of each page (10). The guide says what's screened (the last interaction), how Detect and Enforce differ, and recommends a calibration cycle before enforcing (16 of 20). role is an enum of five values and the flags are booleans, but "messages required unless tools" is stated in prose rather than the schema (12 of 15). Request examples in the docs, and the screen endpoint now documents 400, 401, 422, 429, 500 and 503, the 422, 429 and 503 entries saying whether a retry helps (12 of 15; was 10 for four one-line codes). Versioned path (/v2) and a dated changelog, with a new entry on 26 October 2026 after a gap since 8 June (15; was 13). 25+10+16+12+12+15 = 90.

ShieldGemma 56

Read for a model the owner serves. There is no specification of its own. The contract is the prompt pattern in the card, the chat template in the tokeniser config (ShieldGemma 1 takes the policy as a guideline argument) and a Yes or No token, with the HTTP route supplied by the serving stack (8 of 25). gemma.ai/llms.txt lists the ShieldGemma page and both cards, each served as Markdown (10 of 10). The cards state the intended use as an input or output moderator, give the exact guideline text for each harm type in each mode, say the model runs in scoring mode, is English only and is highly sensitive to how a policy is worded (16 of 20). Fixed harm types and a binary answer, with the policy as free text (10 of 15). A Transformers snippet for each generation and a Keras notebook. No failure modes or error cases are documented (8 of 15). Two named generations with no changelog, and the ai.google.dev page is marked deprecated in favour of gemma.ai (4 of 15). 8 + 10 + 16 + 10 + 8 + 4 = 56.

Lakera Guard (Check Point AI Guardrails) 83

Bearer API keys made on the dashboard. The API reference now says keys carry one of two access scopes chosen at creation, read and write (the default, and the scope of every older key) or read only, where write endpoints return 403. No expiry or rotation documented (24 of 30; was 20). Dashboard roles are User, Admin and No access, and a read-only key can be given to log shippers and dashboards so it can't change policies or projects (14 of 20; was 10). The service is a prompt-attack detector that also screens tool calls and tool results in the same request (15). Every screening request is logged with its prompt and output in the dashboard by default, and logs can be exported to S3 for a SIEM (15). SOC 2 Type II and ISO 27001:2022 on the trust centre per the 1 October check. Check Point publishes a vulnerability disclosure policy covering its products and owned domains, with acknowledgements and no monetary rewards, and posts security advisories in public. No security.txt on lakera.ai or checkpoint.com (both 404 on 4 November), no bug bounty, and no advisory touching AI Guardrails found (15 of 20; was 5 with no disclosure policy found). 24+14+15+15+15 = 83.

ShieldGemma 53

Read as a model the owner runs, on the model lines used for Granite Guardian, with retention and training in place of the read-only line. No key is needed to run it. Google's files need a Hugging Face login and licence acceptance, approved immediately, so there is a token to download but no secret at run time (13 of 30). Self-hosted, so inputs stay on the owner's hardware and nothing reaches Google, though no Google page says so (15 of 20). It reads untrusted text by design. The card says the model is highly sensitive to the policy wording and may behave unpredictably with ambiguous language, and gives no prompt-injection guidance (6 of 15). No call log of its own, though each call returns a score (5 of 15). Weights are safetensors only and load through Transformers classes with no remote code. Google's security.txt is valid and points to its reward programme, and there is no model-specific security page or advisory record (14 of 20). 13 + 15 + 6 + 5 + 14 = 53.

Lakera Guard (Check Point AI Guardrails) 70

status.checkpoint.com still lists Check Point AI Security (Lakera Guard) with Lakera Guard Platform and Lakera Guard API components for the EU, US and APAC (20). Its history feed, read on 4 November, has 25 entries from 27 June to 2 November 2026, and none in the window from 6 August touches the Guard components (the 25 October "Service Down due to incident" page names Check Point Mobile Security only), so a clean record (30). No rate-limit numbers published for Community or Enterprise (0). The API reference now documents 429 as "Too many requests, or the Guard API is temporarily overloaded. The request was not screened. Retry it after a short delay", and 503 with "Retry the request", with no Retry-After header (10 of 15; was 5). No SLA found (0). /v2/guard is GA (10). 20+30+0+10+0+10 = 70.

ShieldGemma 46

Scored on the local-software lines, since the owner runs the model, the reading used for Llama Guard and Granite Guardian. Google's install line is pip install -U transformers[accelerate] with no versions, a Keras Colab notebook covers ShieldGemma 1, and Ollama's own library carries ShieldGemma 1. No Python version or memory figure is stated, and the publisher's files sit behind a login and licence click-through (12 of 20). No CI, tests or evaluation code for the models were found. The technical reports describe the evaluation and nothing reruns it (0 of 25). On the Hub, 8 of 23 threads on the 2B, 2 of 10 on the 9B, none of 8 on the 27B and 9 of 10 on ShieldGemma 2 are open, but every open thread we sampled has a reply from a Google account, the newest on 22 March 2026 (15 of 25). No changelog or version tags. The two generations are named and sit in their own repositories (4 of 15). Released with no preview label (15 of 15). 12 + 0 + 15 + 4 + 15 = 46.

Lakera Guard (Check Point AI Guardrails) 21

Every dated source of the vendor's own was looked for, in the changelog (container releases through 2.0.493 on 2 April 2026, product entries on 10 April, 1 and 8 June and 26 October 2026), no official Guard SDK or MCP package. The newest entry, "Request Latency Percentiles" on 26 October 2026, adds a chart of Guard API request latency to the dashboard's Analytics page and changes nothing the API takes or returns, so it isn't read as an API change. The last API change is still Enhanced Breakdown Response on 8 June 2026, 149 days back (10). One dated entry since 6 August (0). Public changelog and Check Point support, no public issue tracker (8 of 15). No official SDK packages for the Guard API; the red-teaming SDK in the docs is another product (0). Self-hosted container releases on the stable tag every few weeks until 2.0.493 on 2 April 2026 (3 of 10). 10+0+8+0+3 = 21.

ShieldGemma 40

Read as a model, on the lines used for Granite Guardian (recency 30, deprecation 12, churn 8, responsiveness 25, packages 15, health 10). The newest file change is ShieldGemma 2 on 4 April 2025, 19 months before the check, and ShieldGemma 1 last changed in September 2024 (0 of 30). No deprecation policy. Both generations stay downloadable, and the ai.google.dev page was marked deprecated with a pointer to gemma.ai without dates for the models (3 of 12). Two generations in 2024 and 2025, each in its own repository, none replaced in place (6 of 8). Google accounts (members of Google's Gemma organisations on the Hub) answered every thread we sampled on the 2B, 9B and ShieldGemma 2 pages, from August 2024 to 22 March 2026, but leave most threads open (16 of 25). Transformers, Keras and Ollama's library load it (12 of 15). No CI, and Ollama's library build was last updated 2 years ago (3 of 10). 0 + 3 + 6 + 16 + 12 + 3 = 40.

Lakera Guard (Check Point AI Guardrails) 83

The default response is flagged plus a request id, and breakdown, payload and dev_info add detail only when asked (25). The project's policy picks the detectors, and the three flags size each response (20). Six status codes, with 422 marked as not worth retrying unchanged and 429 and 503 as retryable (15 of 20; was 12). A screen has no side effect beyond logging, the reference now says to retry 429 after a short delay and to retry 503, and every call is stored for the dashboard by default (15 of 20; was 12). OpenAI message format in, no official SDK packages for the Guard API (8 of 15). 25+20+15+15+8 = 83.

ShieldGemma 72

Read as a classifier an agent calls. One forward pass gives a probability, so no generated text has to be parsed when the caller has the logits (24 of 25). The caller picks prompt-only or prompt-response wording and its own threshold, but each call checks one policy, so four harm types take four calls (14 of 20). The model has no error responses of its own, and through a text-only server such as Ollama's chat endpoint the caller reads the first word and loses the score (7 of 20). Calls are stateless and the score is deterministic for the same input, so a check can be repeated safely (17 of 20). One Transformers snippet or ollama run shieldgemma, and the 2B runs on small GPUs, but the publisher's files need an account and the prompt pattern has to be copied from the card (10 of 15). 24 + 14 + 7 + 17 + 10 = 72.