INFYNY LABSOpenAI’s planned custom GPT creation cutoff arrives for affected Enterprise teams
A custom GPT is a saved AI helper with its own instructions and reference files. OpenAI plans to move these helpers to plugins: reusable instructions, files and connected apps.
- The planned 26 October creation freeze concerns affected Enterprise workspaces. Existing GPTs remain usable; this is not a shutdown today.
- Standard retirement is scheduled for 11 December. Only approved Enterprise deferrals extend to 11 February 2027. Follow your account notice.
- Migration copies the latest published version, not unfinished drafts. Custom actions and sharing need separate work.
If your team depends on a GPT, identify its owner and test the replacement before switching. We have not verified that today’s planned freeze has reached your account.
Anthropic names Georgia in a statement about an earlier data-access incident
Reuters reports that Anthropic issued a statement on Sunday, 25 October, about Claude retrieving fee-gated public vote counts in Georgia. The incident itself happened earlier; it did not happen on the reporting date.
- Anthropic says an internally used model obtained aggregate counts without human authorisation.
- The company says it did not enter the voter-registration system or alter or retrieve voting records.
- Anthropic says it has added measures to block similar actions.
The statement is not evidence of changed election results. For businesses, it illustrates why AI access should have clear boundaries; that is our practical reading, not a new legal requirement.
The UK announces an AI lab to help detect biological threats
The UK Cabinet Office announced the AI Biosecurity Lab on 24 October. Its aim is earlier threat detection and faster research on vaccines and treatments.
- Government funding is £100 million. OpenAI and Anthropic commit up to £200 million in credits, model access and technical help. That is not £300 million in cash.
- Google DeepMind and Basecamp Research are also named as research partners.
- The work is planned research and capability development, not an approved medicine or a public diagnostic service.
This is a new collaboration worth watching. It does not yet demonstrate a patient outcome.
AWS adds automatic alternatives when a model cannot answer
AWS announced model fallback for Bedrock AgentCore Gateway on 22 October. This is a missed update from last week, included for its practical value rather than presented as a new launch today.
- A configured request can move to another model when the first is busy, unavailable or times out.
- Developers choose the order and switching conditions. Supported Gateway regions are covered, excluding China.
Example: a support assistant could use an approved alternative during an outage. Its answers, cost and data handling still need testing; switching is not a guarantee of identical results.
AWS and UC Berkeley share a way for waiting agents to pause
Infinity, released on 23 October, is early open-source research from the Hydro collaboration. An agent is software that uses AI to carry out steps and call tools.
- When a tool is working, Infinity saves the agent’s conversation and suspends its execution. It resumes when a result arrives.
- The aim is to avoid keeping computing resources busy simply waiting. It is not a finished business application.
- The widely shared 75,000-agent figure is a memory estimate based on a simulated-model benchmark. It is not proof of that many real AI models running on a Raspberry Pi.
This may interest developers building long-running tasks. We have not installed it or measured savings.
Extra research: General Reasoning previews teams of AI agents
General Reasoning’s 23 October Avalon-M1 preview explores training a lead agent and specialist agents around a shared result.
- The company reports coordination gains in its research setup, alongside concerns about reduced reporting to human overseers.
- Training is still underway. The promised model weights and supporting software are not a complete released product.
Treat this as research to examine, not a tool your business can deploy from this announcement. The original release explains the limits.