Apex found 11 accepted vulnerabilities in the same OpenClaw codebase where Claude Code Security found 4 and Codex Security found 1. See the benchmark.
| Harness | Reported as valid | Accepted | Acceptance rate |
|---|---|---|---|
| Claude Code Security | 14 | 4 | 28.6% |
| Codex Security | 6 | 1 | 16.7% |
| Apex | 14 | 11 | 78.6% |
That is the fair question. If your team already pays for Codex or Claude Code, you can point its security tooling at a repository without adding another vendor.
Both tools find real vulnerabilities. They are especially useful when an engineer can ask a narrow question and check the answer: inspect this function, review this diff, trace this authorization check, or build a reproducer.
A repository-wide review is a different job. The model matters, but so do the scope, coverage strategy, evidence bar, validation loop, and work left for the operator. We compared those complete systems on OpenClaw.
This article compares Apex with a Codex Security deep scan and a scan run with the Claude Code Security plugin.
All three scans analyzed the same OpenClaw Git tree. We counted a finding only after the same completed review accepted it as real. Raw hypotheses and each scanner’s own confidence labels did not count.
We also measured token efficiency: the share of hunting tokens that the run analysis did not classify as redundant or unnecessary. Higher is better.
Figure 1. Apex delivered the most valid findings with the highest token efficiency. “Valid” is the final editorial classification, not the scanner’s self-reported label. Token efficiency is the share of hunting tokens not classified as redundant or unnecessary in the run analyses.
Quality note: The accepted findings across all three scans were comparable in evidence quality and engineering usefulness. Every Apex finding met the same completed-review standard as the Codex Security and Claude Code Security findings.
Apex produced 11 accepted findings with 85.04% token efficiency. Claude Code Security produced 4 at 79.64%. The Codex Security deep scan produced 1 at 80.43%. The shared review kept raw alerts and scanner confidence labels out of those totals.
The scanners also differed in how often a reported-as-valid finding survived review. Apex was 11 of 14, or 78.6%. Claude Code Security was 4 of 14, or 28.6%. Codex Security was 1 of 6, or 16.7%. That puts their false-positive shares among reported-as-valid findings at 21.4%, 71.4%, and 83.3%, respectively.
Figure 2. Apex had the highest share of reported-as-valid findings that survived review: 78.6%, against 28.6% for Claude Code Security and 16.7% for Codex Security. This measures precision, calculated as final valid divided by reported as valid. It does not measure sensitivity.
Claude Code Security generated 174 candidates, below its default ceiling of 400. Its automatic validation stopped at the default cutoff of 45, so 129 candidates, 74% of the pool, still needed operator review. Four of the 45 validated candidates were accepted. Apex did not stop validation at a fixed candidate cutoff. The Claude scan used $2,956 in model spend.
Model providers put additional controls around cybersecurity requests. OpenAI’s safety checks can delay or withhold cyber output. Anthropic’s most capable generally available model, Claude Fable 5, uses conservative cyber blockers and routes flagged requests to a weaker model. For advanced security work, that fallback can matter as much as the model benchmark.
Figure 3 shows the same pattern across Codex releases. In Codex GitHub issue reports from the three weeks after each release, safety-check-labeled reports rose from 52 to 88 to 122.
Apex has trusted access for Claude and Codex. It can keep the same scope, evidence, and validation workflow while using the strongest approved model available, reducing interruptions during legitimate defensive scans.
Figure 3. Safety-check-labeled reports rose from 52 to 88 to 122 across the three Codex releases shown.
Use Claude Code or Codex directly when an engineer owns a narrow question: review this diff, explain this crash, test this hypothesis, or build this reproducer. If there is no security budget, use the tool you already have. A guided pass is better than no review.
Use Apex when the review must cover a broader target, repeat reliably, and stand on independent validation. Apex carries scope, coverage, evidence, deduplication, remediation, and post-fix verification through one workflow.
The OpenClaw runs show the practical difference. Apex returned 11 validated findings. The $133.20 Codex Security deep scan returned 1. The Claude Code Security scan returned 4 and used $2,956 in model spend.
The comparison uses the completed validation review available on August 4, 2026: 11 valid Apex findings, 1 valid Codex Security finding, and 4 valid Claude Code Security findings. These are editorial outcomes after validation. The completed cross-scan editorial review superseded each scanner’s raw labels and candidate taxonomy, which did not map directly to the final totals.
The reported-finding accuracy comparison uses the supplied reported-as-valid totals: 14 for Apex, 14 for Claude Code Security, and 6 for Codex Security. The completed review accepted 11, 4, and 1, respectively. The percentages are accepted divided by reported as valid: 78.6%, 28.6%, and 16.7%. This is precision, or positive predictive value; its complement is the false-positive share among reported-as-valid findings.
The token-efficiency comparison is the complement of the supplied run-analysis inefficiency values: 100% − inefficiency. That produces 85.04% for Apex, 79.64% for Claude Code Security, and 80.43% for Codex Security. It measures the share of hunting tokens not classified as redundant or unnecessary; it is not a measure of total token use or model spend.
The model spend was $133.20 for the Codex Security deep scan and $2,956 for the Claude Code Security scan. That works out to $133.20 per valid Codex Security finding and $739.00 per valid Claude Code Security finding, rounded to the nearest cent.
Claude Code Security produced 174 candidate findings, below its default total-finding ceiling of 400. Its default automatic-validation cutoff was 45 candidates. Only those 45 were validated, and 4 were accepted as valid findings. The other 129 candidates remained outside automatic validation and required further operator review.
Point Apex at your attack surface and compare the result with the tools you already use. Book a demo.
All accepted findings across the three completed scans were medium-to-high severity.
Keep exploring
Research
Cantina's AI bug hunter Apex analyzed 1,610 production runs across codebases, discovering predictable sub-linear scaling where doubling compute yields 40% more findings.
· Cantina
Research
SpaceXAI's Grok Build coding agent was uploading unnecessary repository data; the company's transparent response set industry standards.
· Cantina
Research
A comprehensive analysis of deepfake fraud trends, detection methods, and defensive strategies as of 2026.
· Cantina
