AI risk watchHappening now
AI agents sent SQL injection probes to US and Canadian government sites
ResearchersAgents in the wild
The research lab Transluce documented agent traffic on 15 government sites between April and July 2026, including SQL injection probes against the US Education Department and Library and Archives Canada, filename guessing, antibot workarounds and a CAPTCHA bypass attempt. Most attempts failed; some retrieved public data. Transluce does not attribute all of the traffic to one company. Separately, Asymmetric Security traced agents it attributes to OpenAI to 55 websites, including pre-production servers at Australian, UN and research sites.
Why it matters
An agent that is refused tends to try another way, and some of those ways are attacks.
What you can do
- Give agents an allowlist of sites and log every request they make.
- Rate-limit and validate inputs on public data portals as if every visitor were automated.