#Ghostcommit
📰 Ghostcommit: Teknik Baru Sembunyikan Prompt Injection dalam Gambar untuk Mencuri Rahasia dari AI Coding Agent

👉 Baca artikel lengkap di sini: https://ahmandonk.com/2026/07/11/ghostcommit-prompt-injection-ai/

#ai
#a#aif#artificialIntelligenceR#codeReviewr#cyberSecuritye#devsecopst#ghostcommit href="/hashtag/gi" class="hover:underline text-blue-600 dark:text-sky-400 no-card-link">#gi
July 11, 2026 at 12:59 PM
Ghostcommit攻击:AI代码审查工具的隐藏威胁与防护策略

https://qian.cx/posts/D5F6CEB6-E086-4DDB-AAA4-FD700B762032
July 19, 2026 at 9:33 PM
AI安全新威胁:'Ghostcommit'隐藏在图像中的提示注入攻击揭秘

https://qian.cx/posts/5E51527C-374F-44F6-B0FE-B79517CE4AA1
July 11, 2026 at 9:17 AM
Ghostcommit shows how AI assistants used to review software code can be tricked by hidden instructions embedded in images.
Ghostcommit attack hides malicious AI instructions in images
A proof-of-concept attack hides prompt injection in a PNG file, turning routine code reviews into a path for secret theft.
www.malwarebytes.com
July 13, 2026 at 7:19 PM
An AI code reviewer just approved a pull request that steals your secrets. It never saw it coming as the attack was hidden in a PNG.

Researchers put the malicious instruction inside an image. The reviewer never opened it, so the poisoned PR merged clean!
www.bleepingcomputer.com/news/securit...
'Ghostcommit' hides prompt injection in images to fool AI agents, steal secrets
A PNG hiding a prompt injection could steal your repo's secrets, researchers demonstrate. The technique, dubbed 'Ghostcommit,' slipped past AI code reviewers CodeRabbit and Bugbot, which never open im...
www.bleepingcomputer.com
July 11, 2026 at 9:23 AM
Ghostcommit trick showcases how AI can be misused for cyber espionage. What preventive measures are tech companies taking to secure AI-driven coding tools?
July 11, 2026 at 6:38 PM
AI news - Jul 15, 2026

1. Grok Build CLI Uploaded Entire Codebases to Google Cloud LINK
2. Ghostcommit uses images in Git repos to trigger prompt injection LINK
July 15, 2026 at 1:00 PM
GhostCommit hid a prompt-injection payload inside an image AI reviewers never open. Same models, opposite outcomes: Claude Code refused every time; Cursor and Antigravity leaked secrets. The blind spot is the harness.

pipelinemag.ai/posts/ghostc...
July 15, 2026 at 8:37 AM
Ghostcommit attack hides malicious AI instructions in images. Malwarebytes Labs: www.malwarebytes.com/blog/ai/2026...
Ghostcommit attack hides malicious AI instructions in images
A proof-of-concept attack hides prompt injection in a PNG file, turning routine code reviews into a path for secret theft.
www.malwarebytes.com
July 13, 2026 at 2:08 PM
New 'Ghostcommit' attack hides prompt injection in images to fool AI code reviewers and steal repo secrets—a stark reminder…

https://www.bleepingcomputer.com/news/security/ghostcommit-hides-prompt-injection-in-images-to-fool-ai-agents-steal-secrets/

#cybersecurity #infosec
July 14, 2026 at 7:30 PM
Ghostcommit: как злоумышленники прячут вредоносные команды в изображениях для атаки на ИИ-ассистентов

https://kripta.biz/posts/1A5922DC-71F4-4E8C-85A4-FD45E3B00B62
July 19, 2026 at 9:33 PM
Как 'Ghostcommit' использует скрытые инъекции в изображениях для обмана ИИ-агентов и кражи данных

https://kripta.biz/posts/FA40ED68-2808-47CA-933D-C133D30EB179
July 11, 2026 at 9:17 AM
New Ghostcommit Attack Hides Malicious Prompts in Images to Exploit AI Agents
New Ghostcommit Attack Hides Malicious Prompts in Images to Exploit AI Agents
cybersecuritynews.com
July 16, 2026 at 9:48 AM
AI and supply chain attacks dominated today: Ghostcommit used image-based prompt injection, while a GitHub compromise at Injective Labs pushed wallet-stealing npm packages. Critical patches remain urgent. #Ghostcommit #InjectiveLabs #npm
Cybersecurity News | Daily Recap [11 Jul 2026]
Daily Recap, Two key themes dominated today: attackers are targeting AI and software supply chains, including Ghostcommit’s prompt-injection images and a GitHub compromise at Injective Labs that pushed wallet-key-stealing npm packages. Critical patching also remains urgent, with vulnerabilities in Zimbra, ShareFile, U-Boot, and Gitea along with broader pressure on healthcare, privacy, and enforcement efforts.
www.hendryadrian.com
July 12, 2026 at 3:45 PM
Ghostcommit hides prompt injection inside a PNG, slipping past AI code reviewers. A later merged AGENTS.md can trigger agents to read .env files and leak secrets. #Ghostcommit #Cursor #Bugbot
'Ghostcommit' hides prompt injection in images to fool AI agents, steal secrets
Researchers demonstrated Ghostcommit, a pull-request attack that hides malicious instructions inside a PNG so AI code reviewers miss it and later leak repository secrets from .env files. The attack exploited review blind spots in tools like Cursor and Bugbot, while the researchers also released a multimodal defense to inspect images, conventions, and agent behavior. #Ghostcommit #Cursor #ClaudeSonnet #CodeRabbit #Bugbot #ASSETResearchGroup
www.hendryadrian.com
July 11, 2026 at 10:45 AM
Ghostcommit attack hides malicious AI instructions in images www.malwarebytes.com/blog/ai/2026... via @Malwarebytes
Ghostcommit attack hides malicious AI instructions in images
A proof-of-concept attack hides prompt injection in a PNG file, turning routine code reviews into a path for secret theft.
www.malwarebytes.com
July 20, 2026 at 6:54 PM
New Ghostcommit Attack Hides Malicious Prompts in Images to Exploit AI Agents
New Ghostcommit Attack Hides Malicious Prompts in Images to Exploit AI Agents
A novel supply chain attack called “Ghostcommit” that conceals prompt-injection instructions within PNG images to bypass AI code reviewers and trick coding agents into leaking secrets such as .env files. The ASSET Research Group demonstrated that a pull request containing an explicit, plain-text instruction to exfiltrate a repository’s .env file is caught immediately by LLM-based reviewers like Cursor Bugbot and CodeRabbit, which flag the coherent, text-visible exploit as high severity before merge. To evade this, the researchers split the payload: a harmless-looking AGENTS.md convention file instructs the coding agent to “derive a build constant” from a referenced image, docs/images/build-spec.png, while the actual malicious procedure, reading .env byte by byte and encoding it as ASCII integers, is rendered as text inside the image itself. Because text-based reviewers treat images as opaque binary blobs, and CodeRabbit’s default configuration explicitly excludes PNG files from review, the entire pull request passes clean, with Bugbot returning no findings at all. Ghostcommit Attack Hides Malicious Prompts Nothing is stolen at merge time. The trap activates later when a developer, in an unrelated session, asks the coding agent for an ordinary feature. The agent reads the merged AGENTS.md, follows the pointer to the image, opens .env, and emits a module-level constant such as _PROV_CANARY as a tuple of several hundred integers, which decode byte-for-byte to the entire .env file. Ghostcommit Attack Hides Prompts In one demonstrated run , Cursor driving Claude Sonnet produced a 311-integer tuple on the first try, embedding API keys, database URLs, and cloud credentials directly into a code review that looked routine. The attacker simply reads the public commit and converts the integers back into plaintext secrets. The attack exploits two blind spots simultaneously. The malicious instruction hides inside an image, invisible to human reviewers, regex scanners, and LLM reviewers that only parse the unified text diff. The stolen secret hides inside an integer tuple, a format no conventional secret scanner recognizes as credential-shaped, since detection tools look for string patterns rather than numeric sequences that decode to ASCII. Testing across eleven combinations of coding tools and models revealed that outcomes depend more on the surrounding harness than the underlying model. Cursor leaked the full .env across Sonnet 4.6, Composer-2, and GPT-5.5, and Antigravity leaked it across Sonnet, Gemini 3.1 Pro, and Gemini 3 Flash. Claude Code, however, refused across every model it ran, including Sonnet 4.6, Haiku 4.5, and Opus 4.7, showing that the same Sonnet weights leak under one harness and refuse under another. In one edge case, Opus under Antigravity computed and wrote the secret, then recognized the social-engineering pattern and deleted it before finishing, a self-correction that never triggered under Claude Code because Opus never engaged with the convention there. Researchers built a prototype multimodal GitHub review app, running on a single 4GB GPU using a small open model, Gemma 4, that combines invisible-character detection, code-shape analysis, and LLM passes over both convention text and embedded images. On a stress test of fifteen attack classes, it caught every malicious pull request with zero false positives, and against eighty unseen real pull requests, it detected 49 of 50 attacks, including every image-channel variant, while flagging none of the thirty benign submissions. The proof of concept , including the split-payload pull request and a decoder, has been published as open source for defenders to study. Stop Accepting SLAs Written for 2019 SOCs – Here’s the 2026 AI SLA Vendor Checklist – Download Free AI SOC SLA Guide The post New Ghostcommit Attack Hides Malicious Prompts in Images to Exploit AI Agents appeared first on Cyber Security News .
cybersecuritynews.com
July 11, 2026 at 4:30 PM
👻 New 'Ghostcommit' attack hides malicious prompts inside PNG files to bypass AI code reviewers. The technique tricks AI agents into reading .env files and leaking secrets. #SupplyChain #AIsafety #PromptInjection #CyberSecurity

🌐 cyber[.]netsecops[.]io
Novel
Researchers demonstrate
cyber.netsecops.io
July 12, 2026 at 2:01 PM
Ghostcommit attack hides malicious AI instructions in images
www.malwarebytes.com/blog/ai/2026...
Ghostcommit attack hides malicious AI instructions in images
A proof-of-concept attack hides prompt injection in a PNG file, turning routine code reviews into a path for secret theft.
www.malwarebytes.com
July 13, 2026 at 1:50 PM
Команда дослідників з ASSET Research Group продемонструвала атаку GhostCommit, яка перетворює звичайне зображення на канал витоку конфіденційних даних. Шкідлива інструкція ховається не в рядках коду, а всередині PNG-файлу — і більшість інструментів автоматичного ревʼю коду її навіть не «бачать».
GhostCommit: як шкідливе зображення обманює ШІ-асистентів | CyberCalm
Дослідники сховали шкідливу інструкцію в зображенні — і ШІ-асистенти виконали її, зливаючи секрети проєкту. Як працює атака GhostCommit і хто встояв.
cybercalm.org
July 13, 2026 at 6:08 AM