By incident date, November 2024 to October 2026. Amber: three-month average. Dashed: linear trend (-0.11 a month). The bracketed months are provisional.
The stream is the same months as above, split into families: its height is the volume, its bands are the mix. Beside it, each family now against the window before. Hover a band for the month, click it for the incidents; the legend counts open each family.
Movers
The radar is the window's shape against the window before. The heatmap is where each family lands on the severity scale, from research demonstrations to confirmed real-world loss. The bubbles are who is hit by what, and the flow below follows every incident from attack on AI or by AI, through its family, onto its targets.
The shape of the threat
Solid: last 24 months. Dashed: the 24 months before.
How bad, by family
1 research · 2 disclosed · 3 exploited · 4 real loss
Severity 2 here means a flaw that was reported as news. CVEs with no confirmed exploitation are tracked separately below.
Who is hit, by what
An incident can name several targets.
The flow: on AI or by AI, into a family, onto its targets
Ribbon width is the count. The right column counts target mentions, so an incident naming two targets appears in both.
CVEs and advisories for AI frameworks and tools with no confirmed exploitation. They show how much attack surface is being found and patched, not attacks, so none of them are in the counts above. A CVE joins the incidents once CISA or the vendor confirms exploitation.
Most disclosures, by product
Not incidents: a weekly reading of how much web traffic is automated and which AI crawlers send it, from Cloudflare Radar. Background for the counts above, never added to them.
- Bot share is the part of all HTTP requests Cloudflare classifies as automated, over the last seven days.
- AI crawlers are the slice of that bot traffic sent by known AI user agents; their share of all requests is derived from the two, not measured.
- Crawl purpose is what those crawlers declare they fetch pages for: model training, search, or acting on a user’s request.
AI traffic signal
Cloudflare Radar · last 7d
- Googlebot23.6%
- ClaudeBot13.3%
- Meta-ExternalAgent8.3%
- Bingbot7.8%
Source: Cloudflare Radar (CC BY-NC 4.0). Context only; not part of the incident dataset.
The six newest incidents. Each one opens with every source, its classification and related incidents.
- 2026-10-19OpenCode Console Breach Exposes Third-Party AI Provider API Keys3 ExploitedData leak1 src
- 2026-10-19Fake Web3 Accelerator Uses Malicious npm Packages to Distribute Malware3 ExploitedSupply chain1 src
- 2026-10-19Stealthy Poisoning Attacks via Reused KV Cache of Reasoning1 ResearchPoisoning1 src
- 2026-10-19Ignored, Not Refused: Analyzing Indirect Prompt Injection in Agents1 ResearchPrompt inj.1 src
- 2026-10-19ECHO: Response Harmfulness Optimization-Based Jailbreak Attack1 ResearchJailbreak1 src
- 2026-10-19RedDarwin: Evolving Jailbreak Skills against Frontier LLMs1 ResearchJailbreak1 src