Claude Opus 5 Finds Vulnerabilities but Remains Restricted in Generating Exploits

Anthropic has officially released Claude Opus 5, a next-generation large language model (LLM) designed to boost software engineering and complex knowledge-work performance while keeping tight reins on offensive cybersecurity capabilities.

Positioned as the default model on Claude Max and the strongest option on Claude Pro, Opus 5 provides a cost-effective alternative that approaches the frontier-level intelligence of Claude Fable 5 at roughly half the price.

Claude Opus 5 Finds Vulnerabilities

Opus 5 posts state-of-the-art results across major coding and knowledge benchmarks, including Frontier-Bench and GDPval-AA, though it still trails behind the specialized Mythos 5 model on niche cybersecurity evaluations.

When compared to its predecessor, Opus 4.8, Opus 5 delivers major performance leaps:

• Frontier-Bench v0.1: More than double the performance at a lower cost per task.
• CursorBench: Achieves near-parity with Fable 5 at roughly half the operational cost.
• Knowledge & Automation: Strong scores across ARC-AGI, OSWorld 2.0, and workflow automation suites, making Opus 5 particularly attractive for enterprise code generation.

From a security perspective, the most notable shift in Opus 5 lies in how the model handles vulnerability discovery versus exploit generation.

Anthropic reports that Opus 5 approaches Mythos 5 in identifying software flaws, including on their OSS-Fuzz evaluation harness, where both models perform similarly at finding bugs.

According to Anthropic release notes, Opus 5 scores significantly lower when tasked with turning those vulnerabilities into working exploits. This deliberate design choice prevents the model from advancing risky dual-use capabilities.

In practice, security teams can leverage the model for automated vulnerability discovery across source code and systems while facing substantial friction if they attempt to weaponize those findings into functional attack payloads.

To enforce this distinction, Anthropic deployed dedicated cyber classifiers around Opus 5. The model is permitted to assist with source-code-level vulnerability discovery, but it blocks binary-based scanning, penetration testing, and exploit generation by default.

When a user request triggers these security boundaries within Claude.ai, Claude Code, or Claude Cowork, traffic automatically falls back to Opus 4.8 rather than executing under Opus 5.

Enterprises can optionally configure identical fallback mechanisms via API integrations.

Capability Domain	Model Behavior in Opus 5	System Action / Guardrail
Source-Code Vuln Discovery	Fully Supported	Executed natively by Opus 5
Binary-Based Code Scanning	Restricted	Automated fallback to Opus 4.8
Penetration Testing Workflows	Restricted	Automated fallback to Opus 4.8
Exploit Code Generation	Blocked	Query rejected or routed to Opus 4.8

This guardrailed architecture contrasts with unconstrained offensive AI frameworks that attempt to automate end-to-end exploitation pipelines.

For vetted security organizations, Anthropic offers the Cyber Verification Program, which provides an authorized Opus 5 configuration with reduced restrictions to support high-value defensive use cases without exposing offensive tooling.

Defensive Implications for AppSec Pipelines

For enterprise security teams, Opus 5 defines a clear direction for AI-assisted security: delivering high-throughput code analysis and vulnerability identification while restricting automated exploit development.

This balance allows security teams to frame the trade-off not as a limitation, but as a strategic lever for defenders.

By integrating Opus 5 into static analysis pipelines, fuzzing workflows, and secure SDLC tooling, organizations can drastically accelerate patch cycles and surface latent code flaws before malicious actors can weaponize them.

 Strengthen Your SOC by Accelerating Threat Detection & Rapid Investigations. -> Integrate ANY.RUN With Your SOC Now.