AI leaks and early reports
Follow early reports about AI models, assistants and features that may still be in development. ByteForward brings together short editorial summaries, available evidence and the questions each report leaves open. These reports are unconfirmed unless an entry says otherwise. An early interface, screenshot or code reference can show what is being tested without establishing a release date or public availability.
Latest reports
No reports match this search and these filters. Try fewer words, clear search or reset the filters.
Evidence status Still unverified
Paricado labels appeared in a model selector
In April 2026 a Meta AI selector contained Paricado labels for several model types. The entries were inaccessible, so their identity and capabilities were not established. A confirmed release remains unverified.
Evidence status Still unverified
Conway traces pointed to a persistent Claude agent
An April 2026 preview showed Conway, a separate Claude agent environment with extensions and event triggers. Its public release and eventual product name remain unverified.
Evidence status Confirmed later
An early checkout page hinted at a cheaper Pro tier
In February 2026 an early checkout page hinted at a $100 Pro Lite plan. OpenAI introduced a $100 monthly Pro tier in April.
Evidence status Still unverified
Copilot Advisors tested opposing AI viewpoints
In February 2026 a Copilot experiment paired two AI personas on opposite sides of a debate. The visible setup was concrete, while its eventual release remains unverified.
Evidence status Confirmed later
A model comparison selector previewed Model Council
In February 2026 an early selector showed Perplexity combining three models in one query. Model Council launched the next day, confirming the core idea while updating the model lineup.
Evidence status Partly confirmed
A security dashboard appeared in Claude Code tests
A January 2026 preview showed Claude Code organizing security scans and findings in a dashboard. Claude Code Security launched in February with those core capabilities, though the earlier internal names were not officially mapped.
Evidence status Confirmed later
Gemini music tools surfaced before Lyria arrived
Music creation surfaced in Gemini code in January 2026. Google launched the feature with Lyria 3 in February, initially generating short tracks.
Evidence status Confirmed later
Claude widget traces preceded interactive apps
January 2026 Claude tests exposed widget entries beside connected tools. Anthropic released interactive MCP apps five days later, bringing live app interfaces into conversations.
Evidence status Partly confirmed
Anonymous image tests preceded a December upgrade
In December 2025 anonymous image models sparked talk of an OpenAI upgrade. A new image model arrived as GPT Image 1.5, rather than the Image 2 name predicted then.
Questions about AI leaks
What counts as an AI leak
A leak is an early report or product clue that appears before an official announcement. It may involve an interface, screenshot or code reference. The evidence can be incomplete.
Are these reports confirmed
Reports are unconfirmed unless an entry says otherwise. Each summary explains the available evidence and what remains uncertain.
Where can I find the original source
Cards with a verified source link open the original report when selected. Image previews and source credits have their own links. Some historical entries may have no surviving source link.
Why do reports change
New evidence can clarify or contradict an earlier report. Features can also change during testing or never become publicly available.
Does this page cover every leak
Coverage is incomplete. ByteForward selects reports for editorial summaries and cannot capture every clue or independently verify every claim.
How do personal browser reactions work
Reactions need no account and are saved when you choose an icon. The numbers show submitted reactions, and the highlighted icon shows your choice. These totals are not verified unique people. Select the highlighted reaction again to remove it. A random key lets you change choices during the current page visit. Without a previously saved key, reloading or closing the page loses access to those choices, although they remain counted. If you previously chose to remember reactions, the existing cookie can still retrieve your choices until its original expiry or removal through your browser. The current feature does not create reaction cookies or extend their expiry.