THIS DUDE REALLY JUST LEAKED CLAUDE OPUS 4.7 SYSTEM PROMPT!! (link below)
SAFETY
-
Secret Flamingo Unicycle AI Test Results Revealed
By
–
More on my blog, including results from the previously secret "flamingo on a unicycle" test
-
Phasing Out Flawed Evaluation with System Card Caveat
By
–
This is a bad eval that we've been phasing out. Going to add a caveat to the system card to make it clear. More here:
-
Half of world’s largest firms lack critical AI risk framework
By
–
86% report improved productivity': Nearly half of the world's biggest firms lack a critical #AI risk framework — and it’s a dangerous gamble
by Efosa Udinmwen @techradar Learn more: https://
buff.ly/R70ItFH #MachineLearning #ArtificialIntelligence #ML #MI -
Phasing Out MRCR: Shifting Focus from Distraction Tricks to Applied Long Context
By
–
We kept MRCR in the system card for scientific honesty, but we've actually been phasing it out slowly. Two reasons: (1) it's built around stacking distractors to trick the model, which isn't how people actually use long context, and (2) we care more about applied
-
Phasing Out MRCR: Rethinking Long Context Evaluation Methods
By
–
We kept MRCR in the system card for scientific honesty, but we've actually been phasing it out slowly. Two reasons: (1) it's built around stacking distractors to trick the model, which isn't how people actually use long context, and (2) we care more about applied
-
Opus 4.7: Enhanced Long-Task AI With Autonomous Verification
By
–
What makes Opus 4.7 different: in our early testing it handles long-running tasks with more rigor, follows instructions more precisely, and verifies its own outputs before reporting back. Designed for work you can hand off with less oversight.
-
Obliteratus Tool Removes Safety Refusals From Open-Weight LLMs
By
–
Someone built a tool that deletes censorship from any open-weight LLM with a single click.
— AlphaSignal AI (@AlphaSignalAI) 16 avril 2026
Most open-source models ship with built-in refusal.
They reject prompts even when the use case is legitimate.
Research, red-teaming, creative writing.
Obliteratus is a toolkit that… pic.twitter.com/KnNohJirS1Someone built a tool that deletes censorship from any open-weight LLM with a single click. Most open-source models ship with built-in refusal. They reject prompts even when the use case is legitimate. Research, red-teaming, creative writing. Obliteratus is a toolkit that
-
Federated Unlearning: Privacy Improvement or Cybersecurity Risk?
By
–
Does ‘federated unlearning’ in #AI improve #Data #Privacy, or create a new #CyberSecurity risk?
by Abbas Yazdinejad Ann Fitz-Gerald @ConversationUS Learn more: https://
bit.ly/3Qcs31Z #Infosec #IT #Technology -
Agent Evals Drift from Production Reality Standards
By
–
Agent evals are drifting away from production reality. Most benchmarks use clean tasks, well-specified requirements, deterministic metrics, and retrospective curation. Production work is messier, with implicit constraints, fragmented multimodal inputs, undeclared domain