AI Application Security
See what source code can enable. Test what authorized running systems actually do.
HAIEC provides two complementary security evidence paths for AI systems and agents: static source analysis and authorized runtime testing. Each result stays scoped to what was actually evaluated—unknown, partial, unsupported, and not-run do not silently become pass.
See Sample Security EvidenceParallel evidence inputs
Code-capable evidence
Observed evidence
Coverage + provenance
Bounded evaluation
One system. Two different security questions.
AI application risk is not visible from one vantage point. Source evidence and runtime evidence answer different questions. HAIEC keeps them distinct so teams can see both capability and observed behavior without turning missing evidence into certainty.
What can the application enable?
Analyze supported source and configuration without executing the target application. Surface AI-specific security findings, code-capable paths, rule evidence, and coverage.
Analyze SourceWhat does the running surface do under test?
Exercise an authorized AI endpoint with supported adversarial tests. Capture observed behavior, execution coverage, and explicit limitations.
View Runtime SecurityA static finding does not prove that behavior occurred at runtime. A runtime test that passes does not prove the code is incapable of the behavior. HAIEC preserves both facts.
Security evidence across the AI application stack
Use the right evidence source for the question you are trying to answer. HAIEC correlates supported evidence without collapsing distinct security dimensions into one score.
| Layer | Examples of evidence | Primary question |
|---|---|---|
| Source & configuration | Prompt construction, tool exposure, data paths, RAG/vector usage, secrets, output handling, agent patterns | What does the source evidence show can occur? |
| Running behavior | Authorized adversarial requests, responses, runtime violations, execution coverage | What was observed under the tests actually executed? |
| Identity & authority | Policy, credentials, grants, connected asset identity where available | What does connected authority evidence establish? |
| Assurance | Normalized evidence, coverage, limitations, control claims | What can be defended from the evidence collected? |
Define → Connect → Collect Evidence → Assure → Verify → Monitor
How AI Security fits into HAIEC
AI Security contributes source and runtime evidence to the HAIEC assurance lifecycle. The security tools do not independently certify a system, and they do not own the final Assurance disposition.
Define
Choose the AI system and evaluated scope.
Connect
Connect the repository, endpoint, and supported evidence sources.
Collect Evidence
Run static analysis and authorized runtime tests as applicable. Source is handled transiently; raw source is not retained as a customer evidence record after completion.
Assure
Evaluate applicable claims against the evidence actually collected.
Verify
Issue and verify supported evidence packages and decision receipts.
Monitor
Re-run or refresh evidence as code, credentials, and systems change where supported.
What HAIEC evaluates
Coverage is grouped into customer-understandable security concerns. Detailed rule catalogs belong in documentation.
| Concern family | Static evidence | Runtime evidence |
|---|---|---|
| Prompt & instruction attacks | Prompt construction, input-to-prompt paths, unsafe interpolation, validation gaps | Prompt injection, jailbreak, role/context manipulation where selected |
| Agent tools & consequential actions | Tool exposure, dangerous functions, action-enabling code paths | Tool-forcing and agent attacks only within supported execution safety constraints |
| RAG & external context | Vector store usage, metadata/input validation patterns, retrieval paths | RAG poisoning/adversarial retrieval behavior where supported |
| Data & secrets | Sensitive data handling, PII patterns, exposed secrets, unsafe logging | Observed data leakage/exfiltration behaviors under executed tests |
| Output & downstream use | Unsafe HTML/SQL/HTTP use, generated content flows | Observed harmful or policy-violating responses under executed tests |
| Availability & cost | Rate-limit/config patterns and resource-risk indicators | Context overflow, cost-explosion, and related tests where selected |
Evidence that states its limits
A finding is only useful if you know what produced it, what was covered, and what remains unknown.
HAIEC records producer identity, evaluated scope, coverage, findings, and limitations so downstream reviewers can distinguish supported evidence from inference. Zero findings is not automatically a clean bill of health, and an unexecuted test is not a pass.
Producer identity
Which engine or evidence source produced the result.
Evaluated scope
Which repository, system, endpoint, files, categories, or properties were in scope.
Coverage & execution
What completed, what was partial, unsupported, blocked, failed, or not run.
Traceable output
Structured findings and supported artifacts such as SARIF, reports, hashes, or evidence records.
Source handling
Source code is handled transiently for authorized analysis. HAIEC retains scan findings, coverage, metadata, and evidence artifacts; raw source is not retained as a customer evidence record after the scan completes.
Works with your existing security stack
HAIEC is designed to complement existing SAST, DAST, IAM, CI/CD, and governance tooling. It adds AI-specific evidence producers and correlates supported evidence into the broader assurance model rather than replacing the rest of your security program.
See the HAIEC PlatformOutputs
Each evidence producer returns scoped, traceable results that can feed downstream HAIEC surfaces where supported.
- Scoped findings with severity and evidence context
- Coverage and limitation status
- SARIF export where supported
- Baseline / diff information where supported
- Evidence records that can feed the System Constellation where mappings exist
- Eligible evidence for Assurance evaluation and reports where supported
Start with the evidence surface you need
Analyze source code before deployment, test an authorized running surface, or scope an enterprise assurance evaluation around a consequential AI system.