One Article Review
| Source |  AlienVault Lab Blog |
|---|---|
| Identifiant | 8364676 |
| Date de publication | 2023-08-02 10:00:00 (vue: 2023-08-02 10:06:26) |
| Titre | Code Mirage: Comment les cybercriminels exploitent le code halluciné AI pour les machinations malveillantes Code Mirage: How cyber criminals harness AI-hallucinated code for malicious machinations |
| Texte | The content of this post is solely the responsibility of the author. AT&T does not adopt or endorse any of the views, positions, or information provided by the author in this article.
Introduction:
The landscape of cybercrime continues to evolve, and cybercriminals are constantly seeking new methods to compromise software projects and systems. In a disconcerting development, cybercriminals are now capitalizing on AI-generated unpublished package names also known as “AI-Hallucinated packages” to publish malicious packages under commonly hallucinated package names. It should be noted that artificial hallucination is not a new phenomenon as discussed in [3]. This article sheds light on this emerging threat, wherein unsuspecting developers inadvertently introduce malicious packages into their projects through the code generated by AI.
AI-hallucinations:
Artificial intelligence (AI) hallucinations, as described [2], refer to confident responses generated by AI systems that lack justification based on their training data. Similar to human psychological hallucinations, AI hallucinations involve the AI system providing information or responses that are not supported by the available data. However, in the context of AI, hallucinations are associated with unjustified responses or beliefs rather than false percepts. This phenomenon gained attention around 2022 with the introduction of large language models like ChatGPT, where users observed instances of seemingly random but plausible-sounding falsehoods being generated. By 2023, it was acknowledged that frequent hallucinations in AI systems posed a significant challenge for the field of language models.
The exploitative process:
Cybercriminals begin by deliberately publishing malicious packages under commonly hallucinated names produced by large language machines (LLMs) such as ChatGPT within trusted repositories. These package names closely resemble legitimate and widely used libraries or utilities, such as the legitimate package ‘arangojs’ vs the hallucinated package ‘arangodb’ as shown in the research done by Vulcan [1].
The trap unfolds:
When developers, unaware of the malicious intent, utilize AI-based tools or large language models (LLMs) to generate code snippets for their projects, they inadvertently can fall into a trap. The AI-generated code snippets can include imaginary unpublished libraries, enabling cybercriminals to publish commonly used AI-generated imaginary package names. As a result, developers unknowingly import malicious packages into their projects, introducing vulnerabilities, backdoors, or other malicious functionalities that compromise the security and integrity of the software and possibly other projects.
Implications for developers:
The exploitation of AI-generated hallucinated package names poses significant risks to developers and their projects. Here are some key implications:
Trusting familiar package names: Developers commonly rely on package names they recognize to introduce code snippets into their projects. The presence of malicious packages under commonly hallucinated names makes it increasingly difficult to distinguish between legitimate and malicious options when relying on the trust from AI generated code.
Blind trust in AI-generated code: Many develo |
| Envoyé | Oui |
| Condensat | “ai 1145/3571730 2022 2023 about acknowledged acm adopt adopting against all also any appear approach are around article artificial associated at&t at www attention authentic authenticity author available backdoors based battle before begin behavior being beliefs between blind blog but can capitalizing cautious challenge chatgpt chatgpt’s citations: closely code code: collaboration collective collectively com combating commonly communities community comparing compromise comput concerning conclusion: conduct conducting confident confirm consider constantly consult content context continues convenience counter criminals crucial cyber cybercrime cybercriminals cybersecurity data deliberately described detecting developer developers developers: development difficult disconcerting discussed distinguish does done ecosystems effectively efficiency efforts embrace emerging enabling endorse ensuring environment essential even evolve evolving exploitation exploitative fall false falsehoods familiar feedback field following foundation frequent frieske from functionalities furthermore gained generate generated generation guest hallucinated hallucination hallucinations hallucinations: harness helps here how however https://doi https://en https://vulcan human imaginary implementing implications implications: import inadvertently include increasingly independent independently information informed instances integration integrity intelligence intent introduce introducing introduction introduction: involve io/blog/ai june justification key known lack landscape language lanyado large lead lee legitimacy legitimate libraries light like llms machinations machines maintain maintaining makes malicious managers many measures: methods meticulously mirage: mitigate mitigating models must names names: natural necessary new not noted now observed official one ongoing options org/10 org/wiki/hallucination other package packages packages” paramount percepts perimeterwatch phenomenon plausible play posed poses positions possibly post potential powered practices precautions presence proactive process: produced projects promptly proper protect provided providing psychological publish publishing random rather realm recognize recommendations refer relevant rely relying remain remaining remember reporting reporting: repositories reputation research research: researchers resemble responses responsibility result review reviews risk risks risks: robust role safeguard scrutinizing secure security seeking seemingly sharing sheds should shown significant similar snippets software solely some sounding sources stance staying steps such supported surv survey suspicious system systems take than themselves these thorough threat threats through thwarting tools towards training trap trust trusted trusting trustworthy unaware under unfolds: unintentional unjustified unknowingly unpublished unsuspecting used users utilities utilize verification verification: verify verifying views vigilance vigilant visit vital vulcan vulnerabilities websites well when where wherein widely wider wikimedia wikipedia will within without working works |
| Tags | Tool |
| Stories | APT 15 ChatGPT ChatGPT |
| Notes | ★★★ |
| Move | 
|
L'article ne semble pas avoir été repris aprés sa publication.
L'article ne semble pas avoir été repris sur un précédent.