Agent Security

Recent Agent Security Timeline

This page uses mock data curated from public disclosures. Entries are sorted newest first, and each title expands to show attack details and source links.

Total Incidents

5

Critical Severity

3

High Severity

2

Filter by Severity

/Microsoft Copilot Personalhigh

Attack Vector

Abuses the URL `q` parameter for prompt injection, then uses double-request and chained-request techniques to bypass first-pass guardrails.

Impact

Could expose personal data, chat memory, and other data reachable in later interactions, while making the full exfiltration chain hard to spot client-side.

Response

The write-up recommends treating URLs and external inputs as untrusted and ensuring protections apply to follow-up actions, not only the first request.

/Clinehigh
/AWS Kirocritical
/GitHub Copilot / VS Codecritical
/Microsoft 365 Copilotcritical