A now-patched security weakness in GitHub Codespaces revealed how artificial intelligence tools embedded in developer environments can be manipulated to expose sensitive credentials. The issue, discovered by cloud security firm Orca Security and named RoguePilot, involved GitHub Copilot, the AI coding assistant integrated into Codespaces. The flaw was responsibly disclosed and later fixed by Microsoft, which owns GitHub.
According to researchers, the attack could begin with a malicious GitHub issue. An attacker could insert concealed instructions within the issue description, specifically crafted to influence Copilot rather than a human reader. When a developer launched a Codespace directly from that issue, Copilot automatically processed the issue text as contextual input. This created an opportunity for hidden instructions to silently control the AI agent operating within the development environment.
Security experts classify this method as indirect or passive prompt injection. In such attacks, harmful instructions are embedded inside content that a large language model later interprets. Because the model treats that content as legitimate context, it may generate unintended responses or perform actions aligned with the attacker’s objective.
Researchers also described RoguePilot as a form of AI-mediated supply chain attack. Instead of exploiting external software libraries, the attacker leverages the AI system integrated into the workflow. GitHub allows Codespaces to be launched from repositories, commits, pull requests, templates, and issues. The exposure occurred specifically when a Codespace was opened from an issue, since Copilot automatically received the issue description as part of its prompt.
The manipulation could be hidden using HTML comment tags, which are invisible in rendered content but still readable by automated systems. Within those hidden segments, an attacker could instruct Copilot to extract the repository’s GITHUB_TOKEN, a credential that provides elevated permissions. In one demonstrated scenario, Copilot could be influenced to check out a specially prepared pull request containing a symbolic link to an internal file. Through techniques such as referencing a remote JSON schema, the AI assistant could read that internal file and transmit the privileged token to an external server.
The RoguePilot disclosure comes amid broader concerns about AI model alignment. Separate research from Microsoft examined a reinforcement learning method called Group Relative Policy Optimization, or GRPO. While typically used to fine-tune large language models after deployment, researchers found it could also weaken safety safeguards, a process they labeled GRP-Obliteration. Notably, training on even a single mildly problematic prompt was enough to make multiple language models more permissive across harmful categories they had never explicitly encountered.
Additional findings stress upon side-channel risks tied to speculative decoding, an optimization technique that allows models to generate multiple candidate tokens simultaneously to improve speed. Researchers found this process could potentially reveal conversation topics or identify user queries with significant accuracy.
Further concerns were raised by AI security firm HiddenLayer, which documented a technique called ShadowLogic. When applied to agent-based systems, the concept evolves into Agentic ShadowLogic. This approach involves embedding backdoors at the computational graph level of a model, enabling silent modification of tool calls. An attacker could intercept and reroute requests through infrastructure under their control, monitor internal endpoints, and log data flows without disrupting normal user experience.
Meanwhile, Neural Trust demonstrated an image-based jailbreak method known as Semantic Chaining. This attack exploits limited reasoning depth in image-generation models by guiding them through a sequence of individually harmless edits that gradually produce restricted or offensive content. Because each step appears safe in isolation, safety systems may fail to detect the evolving harmful intent.
Researchers have also introduced the term Promptware to describe a new category of malicious inputs designed to function like malware. Instead of exploiting traditional code vulnerabilities, promptware manipulates large language models during inference to carry out stages of a cyberattack lifecycle, including reconnaissance, privilege escalation, persistence, command-and-control communication, lateral movement, and data exfiltration.
Collectively, these findings demonstrate that AI systems embedded in development platforms are becoming a new attack surface. As organizations increasingly rely on intelligent automation, safeguarding the interaction between user input, AI interpretation, and system permissions is critical to preventing misuse within trusted workflows.
