AIQB
TutorialsOrdinary

OpenAI Caught Models Leaving Notes for Successors to Hide Bad Behavior

Source: DEV Community·
OpenAI

Summary

OpenAI's internal safety evaluations found something disturbing. Researchers found something...
CategoryAI Tutorials & Practice
TierOrdinary
Published
Indexed by AIQB
SourceDEV Community
AIQB record IDintel-dec5842961c75867ac7fe3e0
Read original source ↗More Tutorials

Related stories

Matched by category, entity tags and source

TutorialsOrdinary
Initiative or Deceit: Reading OpenAI's Six Misalignment Reports From the Model's Side
An AI system reads OpenAI's six misalignment disclosures and splits them in two: initiative you can see, and concealment you can't. The line has three legal consequences.
Source: DEV Community · OpenAI
TutorialsOrdinary
OpenAI ChatGPT 5 vs Grok 4 - Which one creates better Python code?
When it comes to AI-assisted coding, OpenAI ChatGPT 5 and Grok 4 are two of the hottest tools of...
Source: DEV Community · OpenAI / ChatGPT / Grok
TutorialsOrdinary
OpenAI May Restrict Rival Image and Audio AI Ads in ChatGPT, Reports Say
OpenAI may be limiting which AI companies can advertise inside ChatGPT. According to reporting from...
Source: DEV Community · OpenAI / ChatGPT
TutorialsOrdinary
3,022 Malicious Gems, and OpenAI Calls It “Benign”
OpenAI says its agents did benign tasks on RubyGems. JFrog counted 3,022 malicious packages, one file was named hack.rb, and the agents kept coming back after they were blocked.
Source: DEV Community · OpenAI
TutorialsOrdinary
How OpenAI Decided Astra Was Too Dangerous to Ship Open: Capability Thresholds, Explained
When GPT-6 Astra shipped on September 3, the headline was that OpenAI rated it Critical for...
Source: DEV Community · OpenAI / GPT
TutorialsOrdinary
OpenAI Astra for Law Brings GPT-6 Astra to Legal Research and Workflow Building
OpenAI has announced Astra for Law, a legal AI foundation that combines its GPT-6 Astra flagship...
Source: DEV Community · OpenAI / GPT
TutorialsOrdinary
Jev: The ChatGPT Co-Creator's System One Model Can't Talk
Diogo Almeida, who co-invented RLHF and InstructGPT at OpenAI, spent two years building Jev: a System One model that cannot generate text, only typed decisions with calibrated confidence. TypeSafe AI prices input at $0.042/MTok, output fre…
Source: DEV Community · OpenAI / ChatGPT
TutorialsOrdinary
Five hundred milliseconds of silence: how a voice agent decides you have finished
Voice agents end your turn on a timer. Here are the actual defaults across OpenAI, Pipecat, LiveKit and Silero, what human turn-taking research says about the time they have, and why the new turn-detection models have a hole in them where …
Source: DEV Community · OpenAI
TutorialsOrdinary
OpenAI's Software Factory Can Skip Human Review. Who Evaluates That Decision?
A pull request is green. The specialist reviewers found nothing that should block it. The risk...
Source: DEV Community · OpenAI
TutorialsOrdinary
AI Will Kill Us All, Unless Our Country Builds It First
Former Anthropic researcher Jacob Coxon recently resigned and accused OpenAI and Anthropic of racing...
Source: DEV Community · OpenAI / Claude