K2 Instruct Model Producing Severe Hallucinations During Agent Calls

Forum
1

Hi Groq Team,

We are testing our AI agent using the K2 Instruct model and are encountering major hallucinations that appeared suddenly without any code, prompt, or config changes on our side.

During a normal customer call flow, the model injected unrelated technical metadata, such as:

  • TCP / HTTPS references

  • IPv6 strings

  • “Wikipedia.org”

  • ASN numbers

  • XML-like <datum>, <citation>, <content> tags

  • Random malformed characters like “बहुत b>”

Below is the exact hallucination portion from the transcript:

Agent: बहुत b>
- TCP: en.wikipedia.org.443 (https)
- IPv6: ***.***.c001
- Wikipedia.org
- ASN: 1984
</datum> … <citation> … <content> …

(None of this exists in our prompts or system instructions.)
This is the transciption:

Screenshot 2025-12-13 004910
Screenshot 2025-12-13 004910955×651 9.76 KB

The rest of the conversation was normal, making this even more unexpected.

Please help us resolving this issue

Thanks,

1 Like
2

Hi, thanks for flagging this, reported to engineering team