Perplexity AI Updates: November 3, 2026

1. Perplexity Releases BrowseSafe++, a Prompt Injection Benchmark Built From Injections Comet Saw on Real Pages

Perplexity released BrowseSafe++, an open benchmark of 38,675 attacked and clean page pairs for testing prompt injection detectors and browser agents. Most pairs (32,187) are built around injections that Comet’s detector flagged on real websites between September 2025 and September 2026, out of 1.15 million flagged, with page content rewritten for privacy and the attack kept verbatim; each page comes as HTML, Markdown, an accessibility tree and a screenshot, 24.4% of injections are in a different language from their page, and more than four in ten push the agent toward an action the user never asked for. Perplexity also trained BrowseSafe-Guard, a pair of 0.8B detectors fine-tuned from Qwen3.5-0.8B (one for text views, one for screenshots), which it reports at about 95% F1 across text views versus 18 prompted models, flagging about one clean page in 20 at an estimated $0.18 per 1,000 pages on a rented H200; the post does not say whether the Guard weights are released. Two results matter for agent builders: every model tested still followed some injections it had correctly flagged when later doing the task (from 1.64% of flagged pages for GPT-6 Astra to 24.73% for Llama 3.1 8B), and in the BrowseSafe-Adapt environment, where an attacker model rewrites failed injections up to 64 times, attacks succeeded on at least 79% of tasks against five of seven agents, and on 84.2% of tasks against a Gemini 3.5 Flash-Lite agent even with Guard screening each page first. The dataset is gated on Hugging Face, the SDK installs from GitHub, and a public leaderboard accepts one scored submission per account per week. Source

2. Perplexity’s Analytics API Stops Returning the Credit Source Breakdown for Computer Credit Usage

Perplexity added a November entry to its API changelog: for dataset=credit_usage, responses from /v1/analytics/computer/usage and /v2/analytics/computer/usage currently omit the Credit Source breakdown, while total credit usage and the Model breakdown remain available. Integrations that expect a fixed set of breakdowns should read the returned categories list, use count for total credits and by_categories.Model for the model breakdown when it is returned, and should not treat an omitted breakdown as zero usage. Billing or reporting jobs that parse these responses with a fixed schema are the ones most likely to break. Source