Cisco AI Defense Inspection API vs ShieldGemma

Cisco AI Defense Inspection API 0

No x402, MPP or L402 in the developer docs or the Offer Description (0). No price is published. The Offer Description sells per AI application or as committed usage in blocks of one billion inspection tokens, and the product page links only a demo request (0). No free tier or trial for the Inspection API was found. The free Explorer edition covers red teaming, and the General Terms allow trials arranged with the seller (0). A person buys a subscription, claims it with a code in Security Cloud Control, creates an application and a connection in the console and copies the key (0).

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.

Cisco AI Defense Inspection API 90

Read with the hosted lines, for the Inspection API. status.security.cisco.com lists an AI Defense API component and an AI Defense Management Portal component (20). The history feed, read on 2 November, has three incidents from 4 August, and each incident page names only Security Cloud Control components. Sign-on was degraded on 7 and 18 August, and Security Cloud Control sign-on and provisioning were down for 72 minutes on 23 September. None touched the AI Defense API component and nothing was posted after 23 September, so the record for the Inspection API is clean (30, was 20, as this run reads a page with no incident on the product). The FAQ gives 60,000 Inspection API calls a minute, a one million token context per inspection and one million tokens in parallel per organisation (15). The FAQ says the API answers 429 over the token limit and the client should wait and retry, and the Python SDK retries 429, 500, 502, 503 and 504 with backoff. No Retry-After header is documented (10 of 15). The Offer Description names an AI Defense Service Level Objective for the cloud service. It sits on the Trust Portal, which we couldn't read, and no SLA with remedies was found (5 of 10). The API is v1.0.0 since 28 February 2025 with no beta label (10). 20+30+15+10+5+10 = 90.

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.

Cisco AI Defense Inspection API 80

The newest release note is AI Defense 2026.10.1 of 7 October 2026, and cisco-aidefense-sdk 2.3.0 was tagged on 28 October 2026 (its changelog heading says 22 October) (30). Ten dated product releases from 5 August to 7 October 2026 and SDK 2.2.0 and 2.3.0 in the 90 days to 2 November (20). Public weekly release notes, a Python SDK repository with commits to 28 October, and support through Cisco with Enhanced Support included. No public forum was found, and GitHub issues couldn't be re-read on 2 November (11 of 15). Python is the only official SDK (12 of 15). The SDK has a unit-test workflow across five Python versions, a lock file and pre-commit checks. CI results were not read (7 of 10). 30+20+11+12+7 = 80.

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.

Cisco AI Defense Inspection API 81

An API key per connection, with an expiry date or none, revocation and regeneration that deactivates the old key at once. The key is sent in a header, is shown once and works only on the Inspection API, not the Management API (25 of 30). The key can only inspect. Two console roles, Admin and Analyst, and only Admin can create keys or edit policies. No finer scopes on a key (15 of 20). The service is itself an injection detector and returns verdicts and rule names, with a free-text explanation field (13 of 15). Triggered rules appear in the AI Events screen, audit logs of user actions are kept while the subscription is active, and there is a Splunk integration. Whether clean inspections are logged per call was not found (12 of 15). www.cisco.com has a signed security.txt valid to 1 January 2027 naming psirt@cisco.com, a vulnerability policy and a CSAF directory, and the SDK's SECURITY.md takes reports through GitHub private reporting. None of the 219 advisories in the 2026 CSAF index covers AI Defense. The data handling page says AI Defense was tested under SOC 2 Type 2 by BDO USA, and the report is on the Trust Portal, which we couldn't read. No bug bounty found (16 of 20).

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.

Cisco AI Defense Inspection API 74

Closed service under Cisco's General Terms, version 6.0 of 10 September 2025, and the AI Defense Offer Description, version 2.1 of 23 September 2026. The Python SDK is Apache-2.0 (17 of 30). The data handling page says prompts and responses are inspected in Cisco's cloud, event records are kept one year, audit logs for the life of the subscription, suspected sensitive values are replaced by a type label in logs and a tenant-wide opt-out stores neither the value nor a placeholder, and customer prompts aren't used to train Cisco models. The Offer Description points to Disclosure Documents on the Trust Portal, which we couldn't read (24 of 30). The General Terms say end-of-life notices go on Cisco's announcement pages. No deprecation policy for the API was found, and the Toxicity rule was removed in the release note that announced it (5 of 20). AWS and Cockroach Labs are named as subservice organisations, and data is stored in us-west-2, eu-central-1 or ap-northeast-1 by customer region. A full sub-processor list was not read (16 of 20).

ShieldGemma 54

The weights are under the Gemma Terms of Use, not an OSI licence. They carry distribution conditions, the Prohibited Use Policy and a clause that lets Google restrict usage remotely where it believes the terms are breached. The terms are dated (1 April 2026) with earlier versions archived (15 of 30). Inputs stay on the owner's hardware, so no privacy statement applies. ShieldGemma 1 was fine-tuned on synthetic internal data and public datasets described in a technical report, and ShieldGemma 2 on a WebLI subset and synthetic images with CSAM filtering. No data is released (20 of 30). No deprecation policy or dated model retirements (5 of 20). The weights carry no telemetry of their own, with no statement either way, and downloads go through Hugging Face, Kaggle or Ollama (10 of 20). 15 + 20 + 5 + 10 = 50.

Cisco AI Defense Inspection API 67

A response is one object with is_safe, action, classifications, severity, rules, processed_rules, explanation, event_id and detected_pii, per the SDK's models. No field selection (18 of 25). There are no lists to page. The policy on the connection or enabled_rules in the call picks the rules, and entity lists for PII, PCI and PHI can be narrowed (12 of 20). Errors are JSON with code, message and details, and the docs list 400 messages for chat and HTTP inspection. A keyless request returned 401 with missing api key on 8 October 2026 (14 of 20). An inspection changes nothing but the event log and the token count, client_transaction_id traces a call, and the SDK retries. No idempotency key (14 of 20). messages is the only required field when the connection has a policy. Python is the only official SDK, with a LangChain middleware, a Google ADK plugin and a Go CLI beside it (9 of 15).

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.