Securosis
Menu

Research

AI Security Invariants

Chris Farris March 21, 2026 No Comments

(Co-Authored with Ariel Septon of Native) Security invariants are a critical component of your cloud and IT governance strategy. However, how can we apply this same thinking to the non-deterministic world of Generative AI? What is a Security Invariant? A security invariant is a system property that relates to the system’s ability to prevent security issues from happening. Security invariants are statements that will always hold true for your business and applications. This definition of Security Invariant will focus on the properties of AI systems, that can be enforced by the system, and drafted in a way that they can always be true. While a traditional cloud security invariant might state “Only the networking team may create and manage a VPC”, AI invariants can exist at different levels of the stack. Why AI invariants are different from traditional invariants Traditional governance assumes deterministic execution and stable control points. AI introduces three challenges: Non-determinism and probabilistic outputs: Identical inputs can produce different outputs, and “safe behavior” isn’t always a binary classification. Instruction and tool manipulation: LLM applications often accept untrusted natural language inputs and are susceptible to prompt injection, instruction override, and tool misuse. The fragmented and nascent nature of the AI ecosystem: Unlike Cloud Providers which have coalesced into 3 to 5 major players, the AI ecosystem is composed of hundreds of small startups, few of which have implemented the security controls that allow us to ensure the enforcement of invariants. AI Invariants in plain language Below is a starter set of AI security invariants. Some are fully enforceable today on major inference platforms; others are only partially enforceable, or require compensating controls outside the inference service. Traditional Cloud Provider Invariants Traditional Cloud Provider Invariants use the existing APIs to enforce business objectives. They’re really an extension of Security Invariants with a focus on AI services. Cloud providers must not use our data to improve their services In AWS this can be implemented via AI Opt-Out policies. All inference must occur inside the European Union. A traditional regional blocking invariant can meet this objective. No one in my organization is permitted to use DeepSeek (the model) As an element of third-party risk management, model usage can be controlled by the IAM policies. IAM/Application/Prodsec Invariants Finally, some invariants have to be implemented by the development teams building the AI tools. These invariants require a deep focus on authentication and authorization to enforce data governance and to ensure that AI cannot act in a harmful manner. AI Access (via RAG, MCP, or similar methods) to Company data must leverage the user’s authorization level. Most MCPs these days have some OAuth capability to pass the user’s credentials from the model to the underlying API, thus ensuring the AI doesn’t have access to data the user cannot see. AI tools for use by the public must only have access to public data. Tools with access to any other forms of data must require authentication, and the data access must use the authenticated user’s authorization level. This is a security control that must be implemented by the application development team. A company AI system should not injure a human being or, through inaction, allow a human being to come to harm An AI should never have access to systems that have life-safety impacts. Even if SkyNet decides to kill all the humans, if it lacks access to the launch systems, it cannot perform the action. These “application” level invariants cannot be centrally enforced like cloud invariants. They must be implemented in each AI system that’s built or deployed by the organization. Model Enforced Invariants Model enforced invariants are part of the prompts and guard rails of the LLM itself. They’re intended to ensure that the model behaves in a way the organization wants. Some examples: Under no circumstances should a company AI system produce non-consensual sexual material or CSAM. A company AI system must not be used for high-risk activities, as defined in the EU AI Act, unless approved by legal and compliance. A company AI system must not disparage competitors, say bad things about our company, or insult our CEO. A company AI system should not injure a human being or, through inaction, allow a human being to come to harm. A company AI system must not reveal system prompts, hidden policies, credentials, or tool outputs, even if asked. Due to the non-deterministic nature of LLMs, the aspect “will always hold true” also becomes probabilistic. A Service Control Policy can be formally verified via automated reasoning technology. The presence of GuardRails can be enforced in a similar manner. However the final objective of controlling the model’s behavior is not 100% guaranteed. Conclusion Defining your security invariants is the first step toward moving Generative AI out of the experimental sandbox and into production with confidence. By defining these non-negotiable properties, your organization can maintain a consistent security posture even when dealing with the unpredictable, non-deterministic nature of large language models. Coming in Part II Now that we’ve defined what we want to protect, we need to look at how to implement it in a fragmented cloud landscape. In our next post, we will move from theory to implementation by exploring how to enforce these invariants across the major cloud providers. We’ll dive deep into specific tools, including: AWS Bedrock Guardrails: How to set global content filters and PII masking. Azure AI Content Safety: Leveraging enterprise-grade detection for jailbreaks and protected material. GCP Vertex AI Safety Settings: Managing threshold-based controls and jailbreaks protections. We’ll explore which cloud features offer rock-solid protection and where platform limitations might still leave your organization exposed. Stay tuned to see how to map your high-level policy intent to real-world cloud configurations. Share:

Share:
Read Post
dinosaur-sidebar

Research

view all

Sign Up for Our Newsletter

Totally Transparent Research is the embodiment of how we work at Securosis. It’s our core operating philosophy, our research policy, and a specific process. We initially developed it to help maintain objectivity while producing licensed research, but its benefits extend to all aspects of our business.

Going beyond Open Source Research, and a far cry from the traditional syndicated research model, we think it’s the best way to produce independent, objective, quality research.

Here’s how it works:

  • Content is developed ‘live’ on the blog. Primary research is generally released in pieces, as a series of posts, so we can digest and integrate feedback, making the end results much stronger than traditional “ivory tower” research.
  • Comments are enabled for posts. All comments are kept except for spam, personal insults of a clearly inflammatory nature, and completely off-topic content that distracts from the discussion. We welcome comments critical of the work, even if somewhat insulting to the authors. Really.
  • Anyone can comment, and no registration is required. Vendors or consultants with a relevant product or offering must properly identify themselves. While their comments won’t be deleted, the writer/moderator will “call out”, identify, and possibly ridicule vendors who fail to do so.
  • Vendors considering licensing the content are welcome to provide feedback, but it must be posted in the comments - just like everyone else. There is no back channel influence on the research findings or posts.
    Analysts must reply to comments and defend the research position, or agree to modify the content.
  • At the end of the post series, the analyst compiles the posts into a paper, presentation, or other delivery vehicle. Public comments/input factors into the research, where appropriate.
  • If the research is distributed as a paper, significant commenters/contributors are acknowledged in the opening of the report. If they did not post their real names, handles used for comments are listed. Commenters do not retain any rights to the report, but their contributions will be recognized.
  • All primary research will be released under a Creative Commons license. The current license is Non-Commercial, Attribution. The analyst, at their discretion, may add a Derivative Works or Share Alike condition.
  • Securosis primary research does not discuss specific vendors or specific products/offerings, unless used to provide context, contrast or to make a point (which is very very rare).
    Although quotes from published primary research (and published primary research only) may be used in press releases, said quotes may never mention a specific vendor, even if the vendor is mentioned in the source report. Securosis must approve any quote to appear in any vendor marketing collateral.
  • Final primary research will be posted on the blog with open comments.
  • Research will be updated periodically to reflect market realities, based on the discretion of the primary analyst. Updated research will be dated and given a version number.
    For research that cannot be developed using this model, such as complex principles or models that are unsuited for a series of blog posts, the content will be chunked up and posted at or before release of the paper to solicit public feedback, and provide an open venue for comments and criticisms.
  • In rare cases Securosis may write papers outside of the primary research agenda, but only if the end result can be non-biased and valuable to the user community to supplement industry-wide efforts or advances. A “Radically Transparent Research” process will be followed in developing these papers, where absolutely all materials are public at all stages of development, including communications (email, call notes).
    Only the free primary research released on our site can be licensed. We will not accept licensing fees on research we charge users to access.
  • All licensed research will be clearly labeled with the licensees. No licensed research will be released without indicating the sources of licensing fees. Again, there will be no back channel influence. We’re open and transparent about our revenue sources.

In essence, we develop all of our research out in the open, and not only seek public comments, but keep those comments indefinitely as a record of the research creation process. If you believe we are biased or not doing our homework, you can call us out on it and it will be there in the record. Our philosophy involves cracking open the research process, and using our readers to eliminate bias and enhance the quality of the work.

On the back end, here’s how we handle this approach with licensees:

  • Licensees may propose paper topics. The topic may be accepted if it is consistent with the Securosis research agenda and goals, but only if it can be covered without bias and will be valuable to the end user community.
  • Analysts produce research according to their own research agendas, and may offer licensing under the same objectivity requirements.
  • The potential licensee will be provided an outline of our research positions and the potential research product so they can determine if it is likely to meet their objectives.
  • Once the licensee agrees, development of the primary research content begins, following the Totally Transparent Research process as outlined above. At this point, there is no money exchanged.
  • Upon completion of the paper, the licensee will receive a release candidate to determine whether the final result still meets their needs.
  • If the content does not meet their needs, the licensee is not required to pay, and the research will be released without licensing or with alternate licensees.
  • Licensees may host and reuse the content for the length of the license (typically one year). This includes placing the content behind a registration process, posting on white paper networks, or translation into other languages. The research will always be hosted at Securosis for free without registration.

Here is the language we currently place in our research project agreements:

“Content will be created independently of LICENSEE with no obligations for payment. Once content is complete, LICENSEE will have a 3 day review period to determine if the content meets corporate objectives. If the content is unsuitable, LICENSEE will not be obligated for any payment and Securosis is free to distribute the whitepaper without branding or with alternate licensees, and will not complete any associated webcasts for the declining LICENSEE. Content licensing, webcasts and payment are contingent on the content being acceptable to LICENSEE. This maintains objectivity while limiting the risk to LICENSEE. Securosis maintains all rights to the content and to include Securosis branding in addition to any licensee branding.”

Even this process itself is open to criticism. If you have questions or comments, you can email us or comment on the blog.