Neverland – Back-running reward notification
The rewards may have been partially stolen. Vulnerability Details The DuctLock contract currently permits a new reward to be added to the next epoch […]
Let us help secure your AI agents, LLM-based products and mitigating risks in your custom internal use cases.
CTO Dean Rubin, Othentic Labs
Dean Rubin, CTO of Othentic Labs, partnered with Composable Security to conduct a thorough security review of the Rewards V2 smart contract module. The project aimed to verify the robustness of a new rewards distribution mechanism integrated with EigenLayer and ensure secure cross-chain operations across multiple Layer 2 networks.
CEO Amadeo Brands, YieldNest
Amadeo Brands, CEO of YieldNest, partnered with us to evaluate the security of their Max Vault integration with the Kernel protocol on BNB Chain. The goal was to ensure safe yield generation and optimize protocol robustness before launch.
Chief Architect Nick Velloff, Braintrust
Nick Velloff came to us for a security review of Braintrust, a decentralized talent network. The primary objective was to ensure the secure expansion of the Braintrust platform onto the Base network, validate integrations with third-party services such as Coinbase Onramp, and secure the wallet infrastructure used by its users.
Secure AI integrations based on years of Web2 experience
Every Secure AI Integration and AI Agent audit includes thorough threat modeling, automated testing, and careful manual review by our team of experienced security experts. This process helps teams understand risks across their AI workflows, quickly address vulnerabilities, and maintain strong protection for their products and internal systems.
It includes both application-level and AI-specific components, including LLM integrations, agent logic, tool usage, permissions, data flows, prompt handling, and surrounding infrastructure. This not only helps remove vulnerabilities from the system, but also improves overall resilience, reliability, and security quality.


Audit reports provide a clear and detailed record of every identified vulnerability, categorized by severity from Critical to Informational. Each issue includes practical recommendations from security experts with experience in AI systems, LLM integrations, and agent-based workflows.
When a vulnerability is fixed before deployment, it is marked as resolved. If an issue remains unaddressed, we include an explanation of its potential impact, possible exploitation scenarios, and the project team’s reasoning.
An audit report shows that a team values security, reliability, and safe AI adoption. By auditing Secure AI Integrations and AI Agents, organizations can reduce risks related to data leakage, prompt injection, unsafe tool usage, excessive permissions, insecure workflows, and costly implementation mistakes.
Trust us, others have already done it
We understand that blockchain builders demand security that’s transparent, proactive, and reliable.
>
$
38
B (USD)
in TVL held by audited protocols
95
% clients
wants to have their next audit with us
>
60
% audits
reported and fixed Critical/High issues
50
% clients
already had more then one audit with us

![]() | ||
Tailored security solutions to meet the unique needs of your blockchain projects


Secure your project
Verify personalized threats
Get extensively tested
Undergo manual review
Get detailed report

Take part in a workshops
Introduce security by design
Expand the team’s knowledge
Improve security awareness
Get detailed report

Take advantage of expert advice
Gain confidence
Delegate work to professionals
Use a second opinion
Get detailed report
The rewards may have been partially stolen. Vulnerability Details The DuctLock contract currently permits a new reward to be added to the next epoch […]
Today the Security Alliance (SEAL) announced that its Certification program is moving from pilot into live engagements. SEAL Certification Goes Live: Composable Security in […]
While evaluating Cursor IDE, a behavior was found that looked like a security control but did not behave like one under realistic command-execution patterns. […]