Ghostcommit attack hides malicious AI instructions in images


Ghostcommit is a proof of concept that shows how AI assistants used to review software code can be tricked by hidden instructions embedded in images.

The academic ASSET Research Group showed that an attacker can place instructions inside an image file, point to it in an AGENTS.md file, and get an AI coding agent to follow those instructions during a later task.

A pull request is basically a formal “please review and add my changes” request that a developer sends before changes are added to the main version of a software project. Human reviewers and, increasingly, AI coding tools may review the changes before they’re accepted.

Read more…
Source:  MalwareBytes Labs


Sign up for the Cyber Security Review Newsletter
The latest cyber security news and insights delivered right to your inbox


Related:

  • Hackers are hijacking smart building access systems to launch DDoS attacks

    February 2, 2020

    Hackers are actively searching the internet and hijacking smart door/building access control systems, which they are using to launch DDoS attacks, according to firewall company SonicWall. The attacks are targeting Linear eMerge E3, a product of Nortek Security & Control (NSC). Linear eMerge E3 devices fall in the hardware category of “access control systems.” They are ...

  • Abusing DLL Misconfigurations — Using Threat Intelligence to Weaponize R&D

    January 31, 2020

    Dynamic-link library (DLL) side-loading occurs when Windows Side-by-Side (WinSxS) manifests are not explicit about the characteristics of DLLs being loaded by a program. In layman’s terms, DLL side-loading can allow an attacker to trick a program into loading a malicious DLL. If you are interested in learning more about how DLL side-loading works and how we see attackers using ...

  • Emotet Uses Coronavirus Scare in Latest Campaign, Targets Japan

    January 31, 2020

    Threat actors behind the Emotet malware used the novel coronavirus (2019-nCoV) scare as a hook for their spam email campaign against targets in Japan. 2019-nCoV, which is believed to have originated in Wuhan, China, in the past month, has caused hundreds of deaths and thousands of confirmed cases in China alone. The virus has already spread to ...

  • Attacker’s Tactics and Techniques in Unsecured Docker Daemons Revealed

    January 29, 2020

    Between September and December 2019, Unit 42 researchers periodically scanned and collected metadata from Docker hosts exposed to the internet (largely due to inadvertent user errors) and this research reveals some of the tactics and techniques used by attackers in the compromised Docker engines. In total, 1,400 unsecured Docker hosts, 8,673 active containers, and 17,927 ...

  • A Ransomware Prescription for the Healthcare Industry

    January 29, 2020

    To paraphrase Mark Twain, reports of ransomware’s death have been greatly exaggerated. Ransomware attacks resumed with a vengeance last year, despite conjecture by some researchers that CPU mining would overtake ransomware as a leading threat vector. Instead, the ransomware threat is stronger than ever, impacting more than 750 healthcare providers and racking up recovery costs approaching $4 billion. Some healthcare ...

  • Security Analysis of Devices That Support SCPI and VISA Protocols

    January 28, 2020

    When a legacy protocol is connected via Ethernet, and subsequently to the internet, security issues arise. Standard Commands for Programmable Instruments (SCPI) is a legacy protocol that many advanced measurement instruments support. It can be issued via General Purpose Interface Bus (GPIB), Universal Asynchronous Receiver/Transmitter (UART), Universal Serial Bus (USB), or Ethernet. However, it is ...