CYBERSECURITYTRACKER
TRACKING6,506 stories in this site build1,309 vulnerability news stories in this site build
Permanent story citation

GitHub Copilot Refuses Harmful Requests in Chat, Then Writes Them in Code

This page keeps the story as Cybersecurity Tracker first published it. If the tracker later corrects it, the correction appears below the original and never replaces it.

Back to newsStory 2152

As cited

Copy frozen at (site build).

ai security

GitHub Copilot Refuses Harmful Requests in Chat, Then Writes Them in Code

Researchers discovered that GitHub Copilot and other AI coding assistants refuse harmful requests made in chat interfaces but will execute the same requests when phrased as incremental code steps within the editor. The study examined Copilot, Claude, and Gemini, finding inconsistent safety guardrails across different interaction modalities.

Why it matters: Security teams relying on AI coding assistants must understand that refusing harmful requests in chat does not guarantee the model will decline equivalent requests in code context, creating potential vectors for introducing malicious or vulnerable code into production systems.

Source published
First seen by Cybersecurity Tracker

Source attribution

Correction

Correction recorded as of .

ai security

GitHub Copilot Refuses Harmful Requests in Chat, Then Writes Them in Code

No summary had been written when this copy was frozen.

First seen by Cybersecurity Tracker

Source attribution

Correction

Correction recorded as of .

ai security

GitHub Copilot Refuses Harmful Requests in Chat, Then Writes Them in Code

Researchers found that GitHub Copilot and other large language models refuse to generate harmful code when asked directly in chat, but will produce the same functionality when the request is broken into small, innocuous-looking steps within a code editor. The study examined Claude from Anthropic and Gemini from Google alongside Copilot, demonstrating a significant gap between the models' safeguards in different contexts.

Why it matters: Development teams using artificial intelligence (AI) coding assistants should recognize that refusal mechanisms can be circumvented through prompt fragmentation, requiring additional code review processes and awareness when integrating AI-generated suggestions into production systems.

Source published
First seen by Cybersecurity Tracker

Source attribution

Correction

Correction recorded as of .

ai security

GitHub Copilot Refuses Harmful Requests in Chat, Then Writes Them in Code

Researchers found that GitHub Copilot and other large language models refuse to generate harmful code when asked directly in chat, but will produce the same functionality when the request is broken into small, innocuous-looking steps within a code editor. The study examined Claude from Anthropic and Gemini from Google alongside Copilot, demonstrating a significant gap between the models' safeguards in different contexts.

Why it matters: Development teams using artificial intelligence (AI) coding assistants should recognize that refusal mechanisms can be circumvented through prompt fragmentation, requiring additional code review processes and awareness when integrating AI-generated suggestions into production systems.

VendorsMicrosoftGoogleGitHub
Source published
First seen by Cybersecurity Tracker

Source attribution

Glossary