Advanced Security
Cyber-capable models that pull in callers, wrappers, and related repos before deciding a finding is real.
Security Review
Find the vulnerabilities that can actually be exploited.
Legacy SAST matches patterns and leaves you to decide what is real. Baz security agents investigate each change and post only what they can show is reachable.
Four layers of security review, one investigation
Every layer reports into one review, with severity, confidence, and OWASP and CWE mappings.
- NominateModel hypotheses plus SAST-inside signals
- InvestigateCallers, wrappers, and related repos via the Context Broker
- ProveReachable and relevant, or suppressed
- ReportSeverity, confidence, OWASP and CWE
- FixFixer opens a PR that keeps the invariant
SAST-inside
Static checks feed the agent hypotheses to test, not comments to post.
Agent security reviewer
Reviews hooks, skills, plugins, MCP servers, and model SDK code against the OWASP LLM Top 10.
Basic Security Patterns
Secrets, PII in logs, injection, and missing validation on every PR.
Full codebase scan
Scan a whole repository for risk that predates the diff.
Your security posture as code
Write your security invariants in markdown; review and scans enforce them.
From finding to fix
Fixer opens a PR that preserves the invariant. A human signs off.
An audit trail for AppSec
Every review and scan is a session AppSec can inspect.
What it catches that rules don't
From our vulnerability research and real agent code. In each, the vulnerable code looks correct to a rule.
Broken object-level authorization
Code An export route trusts workspaceId from the query string.
Baz finds The caller is authenticated but never proven to own the workspace.
A gate that fails open
Code Middleware skips token checks for an internal URL prefix.
Baz finds Every handler behind it inherits the bypass. One fix at the gate closes them all.
Build-plane provenance
Code A deploy job runs a script from object storage with cloud credentials.
Baz finds The script can change outside review and runs without an integrity check.
Path traversal in context
Code A file export joins a request parameter onto a root path.
Baz finds Checks normalization and symlink handling before calling it exploitable.
Untrusted context reaching an agent
Code Fetched page content lands in an agent's system prompt.
Baz finds Flags the injection path a text search for injection phrases would miss.
Agent supply chain on autopilot
Code A plugin auto-updates from a default branch; a job runs a latest image.
Baz finds Flags the upstream path into your environment and proposes pinning.
Baz compared with legacy SAST
Scanners like Semgrep and Snyk Code match known patterns. Baz investigates the change and treats static signals as one input.
| Legacy SAST (Semgrep, Snyk Code) | Baz Security Review | |
|---|---|---|
| How issues are foundSAST-inside | Rules match known sources, sinks, and unsafe APIs | Agents investigate the change; static matches become hypotheses to test |
| Authorization and tenant isolationVulnerability research | Code that authorizes the wrong scope looks correct to a rule | Reasons about caller, object, and ownership across files and services |
| False positivesAdvanced Security | Rule hits are triaged after the fact | Unreachable or mitigated signals are suppressed before anything posts |
| Evidence in the findingDocs | Rule ID, file, and line | Exploit path, severity, confidence, OWASP and CWE, cross-repo references |
| CI/CD and supply chainModel vs. SAST | Interpolation patterns and known-vulnerable packages | Provenance: mutable artifacts, floating tags, unverified binaries run with credentials |
| Agent codeAgent attack surface | Prompt injection rarely appears as text a rule can match | Dedicated agent security reviewer, mapped to the OWASP LLM Top 10 |
| Org-specific policycyber.md | Custom rules in a rule language | Security invariants in plain markdown your coding agents also read |
| Old risk already in the repoFull codebase scan | Re-runs the same rules | Full scan with the same posture, findings fixed by Fixer PRs |
Keep your scanner for dependencies and compliance. Baz decides which findings matter.
From alert queues to proven findings
Scanner-first AppSecBaz Security Review
A wall of pattern matches to triageFindings posted only with evidence they are reachable
Old risk waits for someone to touch the fileFull codebase scans find it with the same posture
AppSec reviews every merge or becomes the bottleneckDevelopers fix it in the PR while the change is fresh
Findings stop at a reportFixer opens the fix PR
Questions teams ask first
Does Baz replace Semgrep or Snyk?
Not necessarily. Keep a scanner for dependencies and compliance; Baz adds the reasoning rules cannot express.
Does it only review the diff?
No. A full codebase scan walks the whole repository.
Where does our code go?
Baz is SOC 2 Type II and runs as SaaS or self-hosted. Self-hosted, or SaaS with Private Mode, keeps code in your infrastructure.
Stop triaging. Start fixing what's real.
Turn on Advanced Security for a repository and see findings on your next PR, or run a full codebase scan with us on a repository you choose.