Perplexity Makes AI Security Software Free and Open Source to Protect AI Agents

Artificial intelligence company Perplexity is taking a new step toward safer AI development by making its internal security technology, called SPACE, available as an open-source project. The platform is designed to help developers run autonomous AI agents in isolated environments, reducing the risk of these systems accessing sensitive files, credentials, or networks without proper authorization.

The move comes as AI agents become more capable of performing tasks independently, including running code, processing documents, and interacting with online services. Although these capabilities can improve productivity, they also introduce security risks when an AI system behaves unexpectedly or follows malicious instructions.

By opening access to SPACE, Perplexity hopes to give developers and businesses another tool for building safer AI-powered applications.

What Is Perplexity SPACE?

SPACE is a security sandbox designed to isolate AI agents from the systems they interact with. Rather than relying entirely on an AI model to follow safety instructions, the platform adds technical restrictions around its working environment.

A sandbox creates a controlled space where software can perform authorized tasks without automatically receiving unrestricted access to the host computer or wider network.

According to Perplexity, SPACE provides several important capabilities:

• Isolated execution: AI-generated code runs inside a separate environment.
• Network controls: Developers can manage which external services an AI agent can reach.
• Credential protection: Sensitive credentials can be managed outside the agent's working environment.
• Session recovery: Work can be paused and resumed without necessarily losing progress.
• Environment snapshots: Developers can restore earlier states or create separate workspaces for different tasks.

These features are particularly useful for AI agents that need to execute code, manipulate files, or complete lengthy workflows with limited human supervision.

Why AI Agents Need Better Security

AI agents are different from traditional chatbots because they can interact with tools and take actions on a user's behalf. Depending on their permissions, they may access files, execute commands, browse websites, or connect to external services.

This creates potential problems if an agent misunderstands a request, encounters malicious instructions hidden in a webpage, or attempts an action outside its intended permissions.

For example, an AI coding agent might need access to a project folder to modify software. If it receives unrestricted access to the computer, an error or security vulnerability could expose unrelated files or sensitive information.

A sandbox helps reduce this risk by placing limits on what the agent can access.

However, sandboxing is not a complete security solution. Developers must also configure permissions carefully, protect authentication credentials, monitor activity, and keep the underlying infrastructure updated.

Perplexity Says SPACE Passed Security Testing

Perplexity says it tested SPACE against attempts to escape its isolation boundaries. The company reported that nine advanced AI models were unable to escape the sandbox in 118 attempts during its reported evaluation.

The company also described earlier testing that identified weaknesses in network restrictions, which it says were subsequently addressed. These results are encouraging, but they represent the company's reported tests rather than proof that the platform is immune to every possible attack.

Security testing depends on the models, attack methods, configurations, and environments used. Independent testing will be important for understanding how SPACE performs against a broader range of threats.

For businesses considering the technology, reviewing the technical documentation and conducting their own security assessments should remain essential steps.

Why Perplexity Is Making SPACE Open Source

Open-source software allows developers to inspect available code, adapt it to their requirements, and contribute improvements where permitted by the project's license.

Perplexity's decision could help other companies avoid building an entirely new sandbox infrastructure from scratch. Developers can examine the implementation and evaluate whether it fits their own AI applications.

Open development can also encourage security researchers to identify weaknesses and suggest improvements. This is particularly important in AI security, where new agent capabilities create challenges that may not have existed in traditional software systems.

Perplexity has also worked on other security-focused projects, including tools intended to detect malicious instructions and monitor risky AI agent behavior.

CoreWeave, Akamai and Crusoe Support the Initiative

Perplexity has announced partnerships involving cloud infrastructure providers CoreWeave, Akamai, and Crusoe to help make SPACE available to customers.

These partnerships could make the technology easier for organizations to adopt as they develop and deploy AI agents at scale.

For companies running large numbers of AI workloads, secure execution environments are especially important because a single system may process sensitive data or interact with multiple external services.

The availability of SPACE through infrastructure partners may help organizations evaluate the technology without having to build every component themselves. Specific availability and deployment requirements will depend on the provider and implementation.

What SPACE Means for Businesses and Developers

The release of SPACE could be useful for software engineers, AI startups, enterprise IT teams, and organizations experimenting with autonomous AI systems.

Potential applications include:

AI coding assistants: Restrict the environment in which generated code can execute.

Automated data analysis: Allow agents to process spreadsheets and datasets in controlled workspaces.

Business automation: Enable agents to perform approved tasks while limiting access to unrelated systems.

AI research: Give developers a framework for evaluating how agents behave under restricted conditions.

Long-running AI workflows: Support pausing, resuming, and recovering work while maintaining control over the execution environment.

The benefits will depend on the quality of the deployment and how carefully permissions are configured. Open-source availability does not automatically mean that every implementation is secure by default.

Can SPACE Eliminate AI Security Risks?

No single security tool can eliminate every risk associated with autonomous AI systems. Sandboxes can reduce the potential impact of mistakes or malicious behavior, but they must operate alongside other safeguards.

Organizations should combine isolation with least-privilege permissions, secure credential management, network monitoring, software updates, and human approval for sensitive actions.

They should also test whether their configuration can withstand attacks relevant to their own workloads.

AI agents may become more capable over time, making independent security controls increasingly important. A model's internal safety instructions should not be the only barrier preventing it from accessing sensitive resources.

Final Thoughts

Perplexity's decision to open-source SPACE highlights an important development in AI security: as autonomous agents gain more capabilities, the environments in which they operate need stronger safeguards.

By providing developers with tools for isolation, network control, credential protection, and session recovery, Perplexity aims to make it easier to build AI systems with clearly defined boundaries.

The project could become a useful resource for developers and businesses adopting agent-based automation. Its long-term value, however, will depend on independent security testing, continued maintenance, and responsible implementation.

For the wider AI industry, the message is clear: building smarter AI agents is only part of the challenge. Ensuring those agents operate safely is equally important.