← Blog

Google Cloud, NDAs, and what's actually covered.

If you generate content for client work under NDA, two questions matter more than which model looks best in a demo: who's legally on the hook if something goes wrong, and where does your material actually travel. Here's what Google Cloud commits to in writing when you call Vertex AI / the Gemini Enterprise Agent Platform API directly with your own key.

IP indemnification: two separate protections

Google's generative AI indemnification has two parts, laid out in its original "Shared fate" announcement from Google Cloud's VP Legal and VP/CISO:

  • Training data indemnity: covers claims that Google's training data infringed someone's IP.
  • Generated output indemnity: covers claims that what you generated infringes someone's IP.

Google's own summary: "Training data: Google provides a training data indemnity... Generated output: The generated output indemnity applies to allegations that generated output infringes a third party's intellectual property rights."

The condition that matters: this excludes output you deliberately used to infringe someone's rights. That's about intent, not about which safety parameters you configure in your API calls, so a permissive filter setting for legitimate creative work doesn't void it by itself. Google's own announcement confirms the indemnity applies automatically, with no need to amend your existing agreement.

Not every model is covered, and this is the part people miss

Google's official list of indemnified services states the API is covered "used with generally available versions" of specific model families (Gemini, Imagen, Veo, Codey, PaLM). Preview models are excluded. That single word, GA, decides whether you're covered.

ModelStatusIndemnified
Nano Banana Pro (Gemini 3 Pro Image) GA since 28 May 2026 Yes
Nano Banana 2 (Gemini 3.1 Flash Image) GA since 28 May 2026 Yes
Nano Banana 2 Lite (Gemini 3.1 Flash-Lite Image) GA since 1 Jul 2026 Yes
Gemini Omni Flash Public preview Not yet

Sources: Nano Banana 2 and Nano Banana Pro GA announcement, Nano Banana 2 Lite GA / Omni Flash preview announcement.

! If a project is genuinely NDA-sensitive, Nano Banana Pro/2/2 Lite currently give you indemnification and Omni Flash doesn't, until Google moves it to GA. Worth checking cloud.google.com/terms/generative-ai-indemnified-services yourself before you commit to a pipeline, since this list changes.

What about Fal, Magnific, Higgsfield and other aggregators?

Everything above describes the relationship between you and Google, when you have your own Google Cloud project and API key. If you instead go through a platform that resells access to Google's models, Fal.ai, Magnific, Higgsfield, or similar, that platform is Google's customer, not you. Google's indemnification is a contract with them. Whether any of it reaches you depends entirely on what that platform's own terms promise, and in the three we checked, the direction is reversed: you indemnify the platform, not the other way around.

PlatformWhat their own terms actually say
Fal.ai "You will indemnify, defend, and hold harmless Company..." and "Company makes no representation or warranty that... any Output Content will be... non-infringing." (fal.ai/terms)
Higgsfield "You will indemnify and hold Company... harmless from any losses... relating to... Your Content and/or Outputs" and "Company expressly disclaims any and all liability arising from or relating to Outputs... You are solely responsible for conducting any necessary rights clearances." (higgsfield.ai/terms-of-use-agreement, Sections 12 & 13.2)
Magnific "Legal indemnification" is listed as an Enterprise-plan feature on Magnific's own pricing page. It isn't listed under the individual Premium/Premium+/Pro plans. (magnific.com/pricing)

None of this means these platforms are untrustworthy, disclaiming output liability while requiring user indemnification is standard SaaS practice, Google itself does the same thing to its own resellers. It just means the "Google indemnifies generated output" protection discussed above is specific to being Google's direct customer. Route through an aggregator and, unless that aggregator explicitly says otherwise in writing (as Magnific does for its Enterprise tier), you're the one holding the risk.

A separate promise: your data doesn't train the model

This is not the same protection as indemnification, and it applies more broadly. Per Google's own documentation: "Google won't use your data to train or fine-tune any AI/ML models without your prior permission or instruction. This applies to all managed models on Vertex AI, including GA and pre-GA models."

i So Omni Flash, even in preview, still isn't training on your prompts or your reference images. What it lacks is the IP-lawsuit indemnity above, not this data protection.

Retention isn't automatically zero

By default, Google caches inputs to Gemini models for latency. Google's own docs: cached content is "stored for up to 24 hours in the data center where the request was served." Reaching zero data retention requires specific configuration on your end, or a contractual arrangement for eligible enterprise accounts. Don't assume ZDR by default, confirm it for your project.

Security certifications

Per Google's own compliance resource center, Google Cloud holds ISO/IEC 27001, 27017, 27018 and 27701, SOC 1, 2 and 3 reports, PCI DSS, FedRAMP, and (relevant if you're an EU studio) country-specific frameworks including Spain's Esquema Nacional de Seguridad (ENS). Full detail on the SOC 2 audit scope is on Google's SOC 2 compliance page.

How this actually flows when you use Fuze

This part isn't a Google policy, it's how Fuze is built, so it's worth being explicit about it rather than just asserting it.

  • Fuze's backend runs entirely on your own machine, bound to 127.0.0.1. There is no Fuze-operated server in the request path.
  • Image generation calls go straight from that local backend to Google via genai.Client(vertexai=True, ...), using your own API key.
  • If you use the GCS Upload node in Flow, reference images go to your own Cloud Storage bucket (public or private, your credentials) before being passed to Gemini as a gs:// URI, never leaving Google's infrastructure at that point.

Compare that to routing through an aggregator: your prompt and reference material reach a third party's servers first, which then calls Google on its own credentials and relays the result back to you. That's an extra party in the chain, governed by that aggregator's own retention and security policies, not Google's.

i Quick answer: yes, you're covered. Google's Service Specific Terms define who's covered by account and API key, not by which desktop app made the request. Fuze never becomes a party to that relationship, no Google Cloud account of its own routes your traffic, no proxying your key, no visibility into your prompts or outputs. Functionally the same as if you'd written the genai.Client call yourself.

Conclusions

  • Nano Banana Pro, Nano Banana 2 and Nano Banana 2 Lite are GA and covered by Google's IP indemnification. Omni Flash is not, while it stays in preview.
  • That coverage is yours as long as you hold the Google Cloud account and API key, regardless of which client software makes the call, including Fuze.
  • Going through an aggregator changes who Google's customer is. Fal, Higgsfield and (outside its Enterprise tier) Magnific all put the indemnification burden on you, not on them, per their own terms.
  • Training restriction (no training on your data) is broader than indemnification, it covers preview models too.
  • Default caching keeps inputs for up to 24 hours; true zero data retention needs explicit setup.
  • Certifications (ISO 27001/17/18/701, SOC 1/2/3, PCI DSS, ENS in Spain) are real and independently audited, not marketing claims.
  • With Fuze + your own Google Cloud key, the request path is you → Google, with your own GCS bucket as the only optional extra stop. No Fuze server, no aggregator.
! None of this is legal advice. Indemnification terms, GA status, and certification scope can change, and whether this coverage is sufficient for a specific NDA depends on that NDA's actual wording. For anything contractually binding, have your own legal counsel confirm the current terms.