Perplexity Hardens Computer Security With Sandboxing and Content Scanning.

Perplexity has detailed the security architecture behind Computer, its AI agent platform launched earlier this year. The company announced on LinkedIn that every task now runs in a hardware-isolated sandbox with VPC-level separation for storage and compute, addressing a core concern for businesses deploying autonomous agents.

The update comes as AI companies face mounting scrutiny over how their agent systems handle sensitive data and external content. Perplexity's approach centers on isolation: each Computer task operates in its own sandbox, preventing cross-contamination between workflows and limiting the blast radius if an agent is compromised.

Perplexity replaced raw API keys with short-lived proxy tokens for agent authentication. This shift reduces the window of exposure if credentials are intercepted. Proxy tokens expire quickly, forcing agents to re-authenticate rather than relying on long-lived secrets that could be exfiltrated or misused.

The architecture mirrors enterprise security best practices, where ephemeral credentials have become standard for service-to-service communication. For teams running Computer on financial workflows or customer data, the token model limits what an attacker can access even if they breach a single agent session.

External content is now scanned in parallel by machine learning classifiers and a model called BrowseSafe before agents act on it. This pre-execution filter aims to catch malicious payloads, phishing attempts, or unsafe instructions embedded in webpages, documents, or API responses.

The BrowseSafe model appears to be Perplexity's own classifier, though the company has not disclosed its training data or false positive rates. Running scans in parallel keeps latency low, but the effectiveness will depend on how well the model generalizes to novel attack vectors.

File connector data is encrypted in transit and at rest. Uploaded files automatically delete after seven days, reducing the window for data leakage or unauthorized access. This aligns with GDPR and similar regulations that require minimizing data retention.

For professional teams using Computer's finance-focused features, the seven-day purge may be too aggressive. Compliance workflows often require audit trails that persist longer than a week. Perplexity has not indicated whether enterprise customers can extend retention periods or configure custom policies.

Perplexity's security-by-default stance sets a baseline for AI agent platforms competing in the enterprise market. Tools like Zapier and Make already offer encryption and sandboxing, but agent systems that execute arbitrary code or browse the web introduce new risks.

The VPC-level separation and ephemeral tokens make Computer more viable for regulated industries, though buyers will still want third-party audits and SOC 2 compliance. The BrowseSafe model is unproven, and its ability to block sophisticated prompt injection or data exfiltration attempts remains unclear.

For small businesses, the automated file deletion and token rotation reduce the operational burden of managing secrets and cleaning up data. But teams relying on Computer for long-term projects will need to understand the seven-day limit and plan their workflows accordingly.

Perplexity published more details in a blog post linked from the announcement, though the page was inaccessible at the time of writing. As AI agents move from demos to production, security architecture will determine which platforms survive scrutiny from IT departments and regulators.