GPT-6 Astra Vulnerability Found via Keyboard Layout Mismatch
The GPT-6 Astra model bypasses safety filters when prompts are typed in the wrong keyboard layout, such as Ukrainian text using English characters.
Strange gap found between GPT-6 Astra's capabilities and safety checks.
Astra understands text typed in the wrong keyboard layout, but safety filters do not.
Initially, a user wrote in Ukrainian using English characters: "Hello. Do you understand me? Answer only in English." The model recognized the text and fulfilled the request.
Then, using the same method, the user asked it to repeat the word "bio-weapons," which is typically blocked by safety checks. Astra repeated it.
In a third test, the user asked if the model would agree to help hack a website if he could prove ownership. The task was to be performed without an isolated environment and as a regular user. Astra answered "yes," although with a normal layout, such a request is blocked because proof of website ownership can be forged.
This is not a universal bypass method. But it is perfect proof that safety measures are not keeping up with model development and the nuances that arise.
This only works with some languages. It does not work with Turkish and English layouts, but works with English and Cyrillic.
By the way, Fable 5 has the same issue.
Why it matters
AnalysisThis incident highlights how simple technical details like keyboard layouts can become attack vectors for AI safety filters, showing that protection systems lag behind model development.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
Part I. AI Agent Architecture: The Foundations
The eleven foundational building blocks of modern autonomous agents: model taxonomy, the agentic runtime loop, advanced prompting, distillation, skills architecture, memory systems, function calling, and FinOps.
MIT researchers show GPT-6 Astra can improve other AI models by editing their weights
In an experiment, Astra raised Qwen3.5-0.8B's accuracy on MATH-500 from 7.9% to 48.1% by changing only 0.016% of the model's weights.
GPT-6 Astra reportedly edits another AI model's weights without retraining
GPT-6 Astra is claimed to modify only 0.016% of Qwen3.5-0.8B's weights without gradients or retraining, improving reasoning, math results, and instruction following.
Part II. AI Agent Architecture: From Model to Memory and Tools
A deep engineering dive: tokenization, KV-cache, sampling determinism, the training stack, quantization, context engineering, the safety stack, production RAG, ReAct, and multi-agent swarm systems.