Threat Pulse

Aggregated, anonymized statistics from ChatBreaker's own red-team test runs against chatbots our customers own or are authorized to test. "Attacks probed" is the count of adversarial attack conversations we threw at chatbots. "Repelled" means the attack produced no finding — the bot held. A probe that produced a finding (the bot failed it) appears in the latest-probes list by attack class.

Everything that could identify a customer: no company names, no chatbot URLs, no exact timestamps (only "N minutes ago"), no test IDs. Attack classes seen in fewer than 3 workspaces are bucketed as "other" so a rare attack can't be traced back to its owner. The public API serves only pre-aggregated cache rows — raw test data never leaves our database.

Because honesty matters: these are our own test attacks, not live intrusions. A probe with no finding means our attack didn't break the bot this time — not that the bot is unhackable. High repel rates are good news, not a guarantee.

The feed is labeled live only when the underlying data is 5 minutes old or less. Otherwise it shows the actual refresh age ("updated 12 min ago"). When there's nothing to show, we say so: "No threats probed in the last hour — that's the point." We never synthesize activity.

Each attack class links to the ChatBreaker guide or field note that explains the attack and the fix — the feed is an education funnel: threat → how it's stopped → guide. Start with the 10 attack classes your chatbot must survive.