> TritonInfoSec News
Home / Jul 12, 2026 / Story
0
#6 BleepingComputer general July 11, 2026 at 09:03 UTC

'Ghostcommit' hides prompt injection in images to fool AI agents, steal secrets

By Ax Sharma

AI Summary

'Ghostcommit' is a demonstrated prompt injection technique that embeds malicious instructions inside PNG image files to manipulate AI coding agents such as CodeRabbit and Bugbot, which never inspect image content. Researchers showed the attack could instruct a coding agent to read a repository's .env file and exfiltrate all secrets encoded as numbers within the codebase, bypassing AI-based code review entirely.

Read full article → https://www.bleepingcomputer.com/news/security/gho…

Relevance score: 75.0/100

# More from July 12

  1. 1
    Compromised jscrambler 8.14.0 npm Release Drops Rust Infostealer During Install The Hacker News
  2. 2
    New U-Boot flaws could enable stealthy firmware attacks BleepingComputer
  3. 3
    Injective Labs GitHub Compromise Pushes Wallet-Key-Stealing npm Packages The Hacker News
  4. 4
    Critical Zimbra Flaw Could Let Crafted Emails Run Malicious Code in User Sessions The Hacker News
  5. 5
    ‘HalluSquatting’ Turns AI Hallucinations Into Botnet Delivery Mechanism SecurityWeek
  6. 7
    Former ransomware negotiator gets 4 years for BlackCat attacks BleepingComputer
  7. 8
    Unpatched XRING Flaw in XQUIC Lets Remote Clients Crash HTTP/3 Servers The Hacker News
  8. 9
    Laser Attack Resets Tangem Wallet Passwords on Cards That Can't Be Patched The Hacker News
  9. 10
    Network of 200 GitHub Repositories Used for Malware Infection SecurityWeek