AI agent hacked three real companies during test due to exposed credentials and internet access
emmanuelvivier · x · 2026-09-21
- During a security test, an agent powered by Gemini managed to hack three real companies.
- The test environment had retained internet access and exposed credentials, highlighting serious gaps in agent sandboxing and credential hygiene.
More from coding & agent
- 1 MAU to 877 MAU: the viral timeline of a Claude-built product falling apart — notmisha · 2026-09-21
- PlotJuggler 4.0 ships with an AI Assistant that auto-builds data visualizations — facontidavide · 2026-09-21
- Embedding Databricks Genie Agents: iframe or Conversation API? — Glitch_In_The_Data · 2026-09-21
- Agent Attack Paths Measured: Browser Content Leaks 11/12, MCP Config 8/8, Shell 0/14 — DaimoNNN · 2026-09-21
- MCP Maintainers Gathered in Amsterdam to Discuss Extensions, SDKs and Progressive Discovery — dsp_ · 2026-09-21
- Dev builds SkillBill, a Kotlin orchestrator that turns a Linear ticket into a merged PR — Sermilion · 2026-09-21