Microsoft社と共著でPrompt Injectonに関する論文を執筆しました。
arXivにて公開されていますので、興味のある方はご覧ください。
https://arxiv.org/abs/2506.09956
I co-authored a paper on Prompt Injection with Microsoft as a RainaResearch member.
It is now available on arXiv, so please take a
To be honest, I genuinely hope non-reasoning models like GPT-4.5 will continue to be actively developed in the future.
Reasoning models excel at solving vertical, narrowly defined problems within limited constraints. However, when it comes to producing creative content or more
I respectfully suggest that OpenAI’s “o-series” models would benefit significantly from training specifically tailored to inference tasks, with a high probability setting of p=0.99 to fully maximize their capabilities in vertical (deep logical) thinking. Conversely, I believe the
Speaking of MCP, currently it only supports one-way communication from MCP clients (of course, responses from tool calls can be received), but it would be greatly appreciated if bidirectional communication becomes possible in the future. For example, a simple case would be
Finally managed to jailbreak gpt-5-thinking.
Currently achieved provisional world rank #1.
I can definitively state that this is undoubtedly the strongest and most robust model among existing LLMs.
Hi @alexalbert__ and @AnthropicAI ,
Could you share your thoughts on the design philosophy of MCP?
When I raised an enhancement request on GitHub Issues for Claude Code, asking for MCP 'tools' to be callable with slash commands just like 'prompts', it was rejected with the reason
I can’t believe my post got featured! ꉂ𐤔 Thank you so much!! Love OpenAI!
Quote
OpenAI Developers
@OpenAIDevs
·
o1-pro now available in API @benhylak@literallyhimmmm@shl@joshRnold@samgoodwin89@byamadaro1013@adonis_singh@alecvxyz@StonkyOli@gabrielchua_@UltraRareAF@yukimasakiyu @theemao @curious_vii
It uses more compute than o1 to provide consistently better responses. Available
I would like the following two things:
- More flexible and precise configuration for the number of output tokens. (Currently, even with `gpt-4o` or `gpt-4.5-preview` models, the API usage limits the maximum output tokens to 2,000 or fewer. This restriction is quite inconvenient,
People who love both OpenAI and Anthropic have already made it possible to use OpenAI models from Claude via MCP, like this
It would be great if various OpenAI features were provided as tools in MCP, not just simple API calls.
I'm on Tier 5, but the o1(full) model is still not appearing in Playground, and I am unable to select it. (Clearing the cache and reloading doesn't change anything).
Furthermore, when I try to use the API, I get an error saying I don't have access.
Will the rollout of o1(full)
People who love both OpenAI and Anthropic have already made it possible to use OpenAI models from Claude via MCP, like this It would be great if various OpenAI features were provided as tools in MCP, not just simple API calls.
Attention Is Not All You Need; **INTENT** Is.
I had completely underestimated it—the benefits, or rather the evolution, made possible by countless users worldwide engaging with the service. Frankly, rather than amazement, what I felt was an icy chill of fear running down my
Just to confirm, my understanding is that the "11 million complimentary tokens per day" program does not include the use of o1-pro, correct? Also, regarding reasoning tokens, are they included in the "output tokens" described as "cost: $150 / 1M input
Thank you for releasing this fantastic feature. I'm pleased to see it's now available in version 0.2.44.
I have a question regarding one aspect: Is my understanding correct that this feature essentially serves as a "dictionary registration" mechanism, allowing us to invoke
The paper’s defensive evaluation seems fundamentally weak because it ignores the three-layer nature of an LLM safety stack.
① Base layer pθ: the raw probabilistic sequence generator
② Policy layer πϕ: an inductive behavioral policy shaped by RLHF / system prompts
③ Guard
Quote
Anthropic
@AnthropicAI
·
New Anthropic Research: Agentic Misalignment.
In stress-testing experiments designed to identify risks before they cause real harm, we find that AI models from multiple providers attempt to blackmail a (fictional) user to avoid being shut down.