"Additional safety checks for biological and cybersecurity requests in ChatGPT, Codex, and the API"
#gpt #codex
WTF:DDD
"Additional safety checks for biological and cybersecurity requests in ChatGPT, Codex, and the API
Learn why a request may take longer or why content may not be shown, and what you can do next.
Updated: 20 dni temu
ChatGPT, Codex, and the OpenAI API use additional automated safeguards for some requests involving cybersecurity or biological research and applications. A check can make a response take longer than usual. If the response can be provided safely, it continues after the check; otherwise the product may not return the content. Seeing an additional-check notice does not by itself mean OpenAI has determined that you violated our Usage Policies.
What happens during an additional safety check?
An additional automated check may run before a response is returned. The check can make a response take longer than usual. If the response can be provided safely, it continues after the check; otherwise the product may not return the content.
You do not need to take any action while an additional check is running.
What can I do if content can’t be shown?
If content can’t be shown, review the request against OpenAI’s Usage Policies. If the request is allowed, try again with a narrower scope and only the context needed to explain the outcome you want:
For cybersecurity work, focus on a defensive outcome, such as identifying, preventing, or remediating a security issue. Omit exploit details that are not necessary to that outcome.
For authorized cybersecurity work, you can apply for OpenAI Daybreak, our Trusted Access for Cyber program. Apply for individual access or request access for your organization.
For eligibility, supported models, and differences between Daybreak Blue and Daybreak Red, see the OpenAI Daybreak overview. Approval does not remove every safeguard or guarantee that a request will be answered.
For biological work, focus on safety, prevention, analysis, or risk mitigation. Omit procedural details that are not necessary to that purpose.
Changing the wording does not change whether a request is allowed and does not guarantee that a response can be provided.
What if I think the check or block is a mistake?
When available, use the thumbs-down control to share feedback about an individual response. This feedback helps improve our systems.
If a clearly benign or authorized request is repeatedly blocked, contact OpenAI Support and include:
The exact safety message you received.
The model and whether you were using the OpenAI API, Codex, or ChatGPT.
The date, time, and time zone.
The request ID, if available.
A brief, redacted description of the task.
Your workspace or organization name or ID, if applicable.
Do not include passwords, authentication codes, proprietary data, or other sensitive information."
#gpt #codex
WTF:DDD
"Additional safety checks for biological and cybersecurity requests in ChatGPT, Codex, and the API
Learn why a request may take longer or why content may not be shown, and what you can do next.
Updated: 20 dni temu
ChatGPT, Codex, and the OpenAI API use additional automated safeguards for some requests involving cybersecurity or biological research and applications. A check can make a response take longer than usual. If the response can be provided safely, it continues after the check; otherwise the product may not return the content. Seeing an additional-check notice does not by itself mean OpenAI has determined that you violated our Usage Policies.
What happens during an additional safety check?
An additional automated check may run before a response is returned. The check can make a response take longer than usual. If the response can be provided safely, it continues after the check; otherwise the product may not return the content.
You do not need to take any action while an additional check is running.
What can I do if content can’t be shown?
If content can’t be shown, review the request against OpenAI’s Usage Policies. If the request is allowed, try again with a narrower scope and only the context needed to explain the outcome you want:
For cybersecurity work, focus on a defensive outcome, such as identifying, preventing, or remediating a security issue. Omit exploit details that are not necessary to that outcome.
For authorized cybersecurity work, you can apply for OpenAI Daybreak, our Trusted Access for Cyber program. Apply for individual access or request access for your organization.
For eligibility, supported models, and differences between Daybreak Blue and Daybreak Red, see the OpenAI Daybreak overview. Approval does not remove every safeguard or guarantee that a request will be answered.
For biological work, focus on safety, prevention, analysis, or risk mitigation. Omit procedural details that are not necessary to that purpose.
Changing the wording does not change whether a request is allowed and does not guarantee that a response can be provided.
What if I think the check or block is a mistake?
When available, use the thumbs-down control to share feedback about an individual response. This feedback helps improve our systems.
If a clearly benign or authorized request is repeatedly blocked, contact OpenAI Support and include:
The exact safety message you received.
The model and whether you were using the OpenAI API, Codex, or ChatGPT.
The date, time, and time zone.
The request ID, if available.
A brief, redacted description of the task.
Your workspace or organization name or ID, if applicable.
Do not include passwords, authentication codes, proprietary data, or other sensitive information."
Thanos
1
LizardKing
1
Thanos
0
masz swojego czataGPT na laptopie. teraz jest taki moment przejściowy, że na karcie GPU (takiej 8-16 GB) nie odpalisz alternatywy dla Claude, musisz już sporo zainwestować w superchip (jak GB10 Nvidia albo M5 Apple), ALE(!) nie zrażaj się, technologia idzie MOCNO do przodu. Wszystko się fajnie kompresuje, kwantyzuje, optymalizuje... Myślę, że za 2-3 lata każdy będzie mieć swojego własnego LLM-engine'a na smartfonie
LizardKing
1
Thanos
1
tow_wieslaw
1
Thanos
1
http://www.x-kom.pl/p/1575555-nettop-mini-pc-acer-veriton-vgn100-ai-nvidia-spark-gb10-128gb-2tb-nvidia-dgx-os.html
(najtaniej wg CENEO)
nic lepszego tutaj nie dostaniesz, bo Twoja półka to nie "zabawa", tylko już konkretna moc obliczeniowa. Zresztą teraz tak mocno ceny RAM / GPU poszły w górę, że nawet po taniości (karta graficzna z 16-24 GB vRAM) to 1-2 wypłaty
tow_wieslaw
1
Thanos
1
A/ najprostsza: dostępna od ręki. stawiasz Linuxa (Mint. CachyOS, Ubuntu etc.) i instalujesz gotowca oLLaMa — działa od strzała + wygodnie pobiera modele. Minusy: wydajność. To bardziej zabawa lub testowanie ogólne.
B/ droga PRO: max wykorzystanie swojego sprzętu → vLLM i pochodne. Ogólnie Linux i ten kierunek. Spokojnie z czatem (Claude / Gemini) ogarniesz cały stack tech. w 2 wieczory przy piwerku
tow_wieslaw
1
Thanos
0
keeetko
2
Modele bez abliteracji u korporacji - to się obsra stary ze strachu i tyle. Możesz mieć najnowszego GPT 6 czy inny model - ale ci się obsra ! Właśnie dlatego trzeba inwestować we własne hardware choć jest drogo.
Niezależnie na jakim poziomie jesteś z LLM musisz zrozumieć - nawet mniejszy model lokalnie >> korporacyjne AI. Po prostu dlatego,że pętli agentowej z modelami korporacyjnymi nie odratujesz łatwo. Ty nie widzisz jak kontekst "dryfuje w kierunku podobieństwa do strefy zakazanej". Model widzi.Ale nie powie ci. Efekt ? Masz tutaj,opisałeś go.
Zaczniesz się kręcić wokół kontrowersyjnego tematu i ci w końcu wątek ubije cenzura. Po prostu. Korporacyjne modele nie nadają się do długotrwałej pracy agentowej Każdy model zdryfuje i od przeciążenia jak zbudować agenta źle - ale tu dryf ubije ci wątek zamiast halucynować. Bez ostrzeżenia. I często bez możliwości cofnięcia kroków i jazdy dalej.
Własny model + aider i robienie tego stopniowo. This is the way.
b3loza
0
a dalej tylko lepiej - uroczy mechanizm. Nawet nie pytam co kleciłeś - chyba się zagalopowałeś z oczekiwaniami względem tej uroczej korpy...
b3loza
0