AI Risk Assessment Readiness and Security Testing in California

Which California AI laws for SaaS companies actually apply?

California's AI rules sort into two groups. One governs how you use personal information, including AI that decides things about people. The other governs what kind of AI you build or offer to the public. Most SaaS companies live in the first group and touch the second only at the edges.

The first group starts with the CCPA. It covers for profit businesses that do business in California and meet any one of three tests. The tests are $26.625 million or more in gross annual revenue; buying, selling, or sharing the personal information of 100,000 or more California residents or households; or earning 50% or more of revenue from selling or sharing it. Nonprofits and government agencies are generally outside it. If you meet none of those tests, the CPPA rules below do not reach you, although the employment rules still can.

This is a plain English guide with dates and exclusions, not legal advice. Take scope questions to your counsel.

What does CCPA ADMT compliance require, and from when?

The CPPA board adopted these regulations on July 24, 2025, and they took effect January 1, 2026. The part on automated decisionmaking technology, or ADMT, gives businesses until January 1, 2027. It applies when you use ADMT to make a significant decision about a consumer, meaning one that grants or denies financial or lending services, housing, education, employment or contracting work and pay, or healthcare.

ADMT means technology that processes personal information and replaces human decisionmaking, or substantially replaces it. Human review only counts if the reviewer understands the output, weighs it with the other facts, and can change the outcome. Covered businesses owe consumers notice before use, a right to opt out or appeal, and a right to access information about its use.

Advertising to a consumer is not a significant decision. Neither is a feature that drafts replies, summarizes calls, or searches documents, as long as it does not decide one of those outcomes. A hiring tool that ranks candidates without real human review is covered.

When is a California AI risk assessment required?

Since January 1, 2026, a business has to finish a risk assessment before it starts any processing the rules treat as a significant privacy risk. Processing that was already running has until December 31, 2027. The first attestation and a summary go to the CPPA by April 1, 2028, then every April 1.

For AI teams, the last three triggers in the list are the sneaky ones. A feature that scores employee performance, or a plan to train a hiring model on applicant data, can turn an ordinary release into a CCPA risk assessment for AI processing. Processing that hits none of the triggers needs no assessment. A support assistant answering product questions from public docs, with no sensitive data involved, is a typical example.

• Selling or sharing personal information.
• Processing sensitive personal information, with a narrow exception for employee pay, benefits, work authorization, and similar required HR tasks.
• Using ADMT for a significant decision about a consumer.
• Using automated processing to infer traits like ability, work performance, health, or behavior from systematic observation of applicants, employees, contractors, or students, or from anyone's presence at a sensitive location.
• Using personal information to train ADMT for significant decisions, or to train facial recognition, emotion recognition, or identity verification technology.

Who needs a CCPA cybersecurity audit, and does an LLM test count?

The audit duty applies to businesses that earned 50% or more of revenue from selling or sharing personal information, or that had revenue over $26,625,000 and processed the personal information of 250,000 or more consumers, or the sensitive information of 50,000 or more. First audit reports are due April 1, 2028 for businesses with 2026 revenue over $100 million, April 1, 2029 for those between $50 million and $100 million, and April 1, 2030 for everyone else covered. A company below those lines owes no audit under these rules.

The auditor must be an independent, qualified professional who did not build or maintain the program under review. The audit examines your entire security program, so an LLM penetration test is not one, and Datamart does not perform these audits. The test gives that auditor AI findings, fixes, and retest results to examine.

Do SB 53, AB 2013, SB 942, or SB 243 reach your product?

SB 53, the Transparency in Frontier Artificial Intelligence Act, was signed September 29, 2025 and took effect January 1, 2026. It covers developers that trained a model using more than 10²⁶ operations, and most duties fall on large frontier developers with over $500 million in revenue. A SaaS company that calls or integrates someone else's model is not a frontier developer under it.

AB 2013 was signed September 28, 2024 and took effect January 1, 2026. Anyone who designs, produces, or substantially modifies a generative AI system offered to Californians must post a high level summary of its training data. The law does not define substantially modifies, so fine tuning a model and offering it publicly could be caught. AI that only classifies or scores, and generates no content, is generally not generative AI as the law defines it.

SB 942, the California AI Transparency Act, was signed September 19, 2024, and AB 853 moved its start to August 2, 2026. Providers of publicly accessible generative AI systems must offer a free AI detection tool and embed hidden provenance labels in AI generated image, video, and audio. Recent amendments widened who is covered. As signed, the labeling duty is written for image, video, and audio, so ask counsel how it applies to a text only assistant.

SB 243 was signed October 13, 2025 and took effect January 1, 2026. Operators of companion chatbots must tell users they are talking to AI, keep suicide and self harm protocols, and protect minors. Bots used only for customer service, productivity, or technical help are excluded.

How does AI compliance in California reach hiring and HR tools?

Civil Rights Council regulations on automated decision systems in employment took effect October 1, 2025. Using an automated tool that discriminates on a protected trait is unlawful, evidence of bias testing counts in both claims and defenses, and records must be kept for 4 years. SB 947, the No Robo Bosses Act, was signed September 30, 2026 and takes effect July 1, 2027. Employers may not rely solely on an automated system to discipline or fire someone, and they must give written notice when they mostly relied on one.

If you sell HR, scheduling, or performance software, your customers are the employers carrying these duties, so expect them to ask for bias testing evidence and a way to keep a human in the decision. Your own company is an employer too, so they also cover AI you use on your staff. A tool that never touches decisions about workers sits outside both.

Where does technical AI security testing help, and where does it stop?

A risk assessment documents the harms a processing activity could cause and the safeguards against them, and a cybersecurity audit checks whether controls work. Testing shows whether they hold up under pressure.

Datamart does readiness work and technical testing, in the four ways below. We do not perform the independent cybersecurity audit described in the CCPA rules, we do not certify compliance, and we do not give legal advice. Your counsel decides scope, and we supply the engineering evidence.

• Evidence for risk assessments. We test the AI features an assessment names for data leakage, access control bypass, and prompt injection, then record what held.
• Findings that feed a cybersecurity audit. Severity rated findings, reproductions, fixes, and retest results give the independent auditor material to review.
• Testing chatbots and agents. We check that AI disclosures, self harm protocols, and human approval steps survive jailbreaks, and that agents stay inside their permissions.
• Readiness mapping. We list each AI feature, the data it touches, and the rules it might trigger, for counsel to review.