OpenAI започна да пуска GPT-6 Astra — първия модел на компанията, на който тя присвои критично ниво на кибервъзможности. Компанията пише, че с нужните инструменти и достъп моделът може да открива неизвестни досега уязвимости и да разработва начини за експлоатацията им в добре защитени системи без поетапни указания от човек.
OpenAI засили защитите срещу вредни кибердействия, причинени от злоупотреба с модела или от несъгласуваните му действия. За външното внедряване на Astra компанията добави наблюдение на всички заявки на модела към инструменти, за да открива несъгласувано поведение.
Проверка на твърденията:
- OpenAI започна да пуска GPT-6 Astra — първия модел на компанията, на който тя присвои критично ниво на кибервъзможности. (потвърдено от самата публикация: доказателство; «Today, we are releasing GPT‑6 Astra, the most capable model we have ever broadly deployed. Astra is our first model to reach the Critical level of cybersecurity capability under our Preparedness Framework.»)
- Компанията пише, че с нужните инструменти и достъп моделът може да открива неизвестни досега уязвимости и да разработва начини за експлоатацията им в добре защитени системи без поетапни указания от човек. (потвърдено от самата публикация: доказателство; «This means that, with the right tools and access, GPT‑6 Astra can find previously unknown security flaws and develop new ways to exploit them across many well-protected systems without a person guiding each step.»)
- OpenAI засили защитите срещу вредни кибердействия, причинени от злоупотреба с модела или от несъгласуваните му действия. (потвърдено от самата публикация: доказателство; «Accordingly, we significantly strengthened our protections against the model taking harmful cyber actions, whether that’s due to misuse or misalignment.»)
- За външното внедряване на Astra компанията добави наблюдение на всички заявки на модела към инструменти, за да открива несъгласувано поведение. (потвърдено от самата публикация: доказателство; «For these reasons, we have additionally added misalignment monitoring to all tool-using inference involved in our external deployment of Astra, with significant compute cost.»)
Публикации:
- https://wired.com/story/openai-astra-first-ai-model-with-critical-cyber-abilities
- https://techcrunch.com/2026/09/01/open-ais-astra-model-is-on-the-way-and-very-good-at-breaking-into-computer-systems
- https://thenewstack.io/astra-api-safety-stops
- https://techcrunch.com/2026/09/02/openais-new-reasoning-technique-alarms-ai-safety-experts
- https://techcrunch.com/2026/09/03/openai-launches-astra-its-powerful-and-controversial-new-model
- https://thenewstack.io/openai-gpt6-astra-benchmarks
- https://wired.com/story/openai-says-gpt-6-can-use-a-computer-better-than-a-human
- https://cnbc.com/2026/09/03/open-ai-astra-gpt-6-cyber.html
- https://giftarticle.ft.com/giftarticle/actions/redeem/1054f4f0-cac7-479c-a4a7-f95b3906ca4b
- https://thenewstack.io/astra-arc-agi-benchmark
- https://simonwillison.net/2026/Sep/3/gpt6-astra
- https://latent.space/p/astra
- https://latent.space/p/ainews-gpt-6-astra-openais-biggest
- https://thenewstack.io/openai-astra-harness-arc-agi-3
- https://thenewstack.io/gpt6-astra-developer-access-delayed
- https://lesswrong.com/posts/tpLBiFKe62HZu5x7B/openai-s-astra-alignment-claims-are-dubious-and-there-is
- https://producthunt.com/products/openai
- https://lesswrong.com/posts/2GkrxFh8CsgFfRhy5/an-ai-slowdown-is-better-than-a-pause
- https://lesswrong.com/posts/QHoF3tJvryRtmAmMg/claude-mythos-5-1-and-fable-5-1-capabilities
- https://thenewstack.io/astra-reasoning-effort-cost
- https://lesswrong.com/posts/GcYpJXqoaQHTvxLRP/blocking-monitors-are-bad
- https://simonwillison.net/2026/Sep/7/llm
- https://lesswrong.com/posts/HCRs8btkiamtWSNAL/astra-is-hard-to-monitor
- https://lesswrong.com/posts/FsCkkoGsNmPzFKRhg/gpt-6-astra-can-do-a-lot-of-multi-hop-reasoning-without
- https://lesswrong.com/posts/AmFJZyeCgvFjNKgNk/gpt-6-astra-the-system-card-alignment-and-what-comes-next
- https://magazine.sebastianraschka.com/p/gpt-6-astra-looped-transformers-and
- https://infoq.com/news/2026/09/openai-gpt6-astra
- https://twitter.com/aimaddie/status/2098128761388716353
- https://techcrunch.com/2026/09/10/openai-puts-pro-subscriptions-on-hold-due-to-astra-demand
- https://thenewstack.io/openai-slowing-ai-development
- https://wired.com/story/the-ai-slowdown-is-an-antitrust-mess
- https://wired.com/story/heres-how-an-ai-slowdown-could-actually-work
- https://wired.com/story/kernel-panic-ai-vulnerability-explosion
- https://independent.co.uk/news/lawsuit-anthropic-google-openai-lawyers-b3052942.html
- https://apnews.com/article/antitrust-lawsuit-ai-slowdown-anthropic-openai-spacexai-google-960af4308161eaf4ed13c383b0ce1c1b
- https://tomshardware.com/tech-industry/big-tech/anthropic-openai-spacexai-and-google-face-antitrust-lawsuit-for-agreeing-to-slow-ai-development-plaintiffs-say-plan-has-been-in-motion-for-months-before-calls-agreement-self-serving
- https://thenewstack.io/gpt-sol-alignment-gaps
- https://thenewstack.io/openai-prompt-caching-costs
- https://drivingbench.com/
Първоизточници:
- https://openai.com/index/path-to-astra
- https://lesswrong.com/posts/AayZFiHRdiFYWuxyx/why-openai-s-astra-could-make-ai-doom-harder-to-prevent
- https://twitter.com/OpenAI/status/2095527557924082061
- https://developers.openai.com/api/docs/models/gpt-5.6-sol
- https://openai.com/index/how-two-settings-tripled-our-arc-agi-3-scores
- https://epoch.ai/benchmarks/frontiermath-tier-4-v2
- https://benchcad.com/news.html
- https://tbench.ai/news/terminal-bench-science-0-1
- https://openai.com/index/gpt-6-astra
- https://deploymentsafety.openai.com/gpt-6-astra
- https://openai.com/index/safety-overview-gpt-6-astra
- https://openai.com/index/legora-financial-statement-review-with-astra
- https://arcprize.org/arc-agi/3
- https://arcprize.org/blog/astra
- https://openai.com/index/playco-game-prototyping-with-astra
- https://artificialanalysis.ai/articles/benchmarking-gpt-6-astra
- https://x.com/OpenAI/status/2095595741528125780
- https://x.com/OpenAI/status/2095595752815030713
- https://x.com/sama/status/2095600005772104059
- https://x.com/OpenAI/status/2095595757072191802
- https://x.com/OpenAIDevs/status/2095596178117419365
- https://x.com/thsottiaux/status/2095597168816226335
- https://arcprize.org/blog/arc-agi-3-launch
- https://arcprize.org/results/openai-gpt-6-astra
- https://x.com/sama/status/2095678759651438887
- https://x.com/thsottiaux/status/2095651088502591861
- https://vercel.com/changelog/gpt-6-astra-now-available-on-vercel-ai-gateway
- https://twitter.com/OpenAI/status/2095968413646737608
- https://openrouter.ai/openai/gpt-6-astra
- https://deploymentsafety.openai.com/gpt-6-astra/external-evaluation-for-monitorability---uk-aisi
- https://deploymentsafety.openai.com/gpt-6-astra/external-evaluations-for-alignment-uk-aisi
- https://github.com/openai/monitorability-evals
- https://arxiv.org/abs/2512.18311
- https://metr.org/hugging-face-incident-report-aug-2026.pdf
- https://simonwillison.net/2026/Sep/4/astra-pelicans
- https://coderabbit.ai/blog/gpt-6-astra-code-review-evaluation
- https://x.com/thsottiaux
- https://artificialanalysis.ai/models/gpt-6-astra-low
- https://artificialanalysis.ai/models/gpt-5-6-sol-high
- https://developers.openai.com/api/docs/guides/latest-model
- https://dev.to/shinpr/switching-from-gpt-56-sol-to-gpt-6-astra-start-with-medium-effort-25ao
- https://blog.redwoodresearch.org/p/how-will-we-update-about-scheming
- https://openai.com/index/pacing-model-development-cyber-capabilities
- https://github.com/simonw/llm/releases/tag/0.35
- https://deploymentsafety.openai.com/gpt-6-astra/misalignment-monitoring
- https://openai.com/index/an-alien-mind
- https://theinformation.com/articles/secret-technique-behind-openais-astra-model-sparks-security-concerns?rc=tv1dv6
- https://tomekkorbak.com/cot-monitorability-is-a-fragile-opportunity/cot_monitoring.pdf
- https://rohansubramani.github.io/astra-no-cot
- https://deploymentsafety.openai.com/gpt-6-astra/alignment
- https://deploymentsafety.openai.com/gpt-6-astra/cybersecurity---trusted-access-for-cyber
- https://deploymentsafety.openai.com/gpt-6-astra/expert-led-assessments
- https://deploymentsafety.openai.com/gpt-6-astra/external-evaluations-for-cyber-capabilities-irregular
- https://deploymentsafety.openai.com/gpt-6-astra/forecasting-misaligned-behavior-with-deployment-simulation-of-internal-codex-traffic
- https://deploymentsafety.openai.com/gpt-6-astra/safeguards
- https://lesswrong.com/posts/ntKx9YHWCwxSeGbRB/estimating-gpt-6-astra-s-no-cot-time-horizon
- https://artificialanalysis.ai/agents/coding-agents
- https://artificialanalysis.ai/evaluations/artificial-analysis-intelligence-index
- https://theinformation.com/articles/secret-technique-behind-openais-astra-model-sparks-security-concerns
- https://arxiv.org/abs/1807.03819
- https://arxiv.org/abs/2607.22083
- https://openai.com/index/gpt-6-astra-next-generation-work
- https://learn.chatgpt.com/docs/config-file/config-reference
- https://openai.com/daybreak
- https://x.com/thsottiaux/status/2098113585683808624
- https://x.com/thsottiaux/status/2097559315150426222
- https://lesswrong.com/posts/eRmzz8J8Qkzqvzrgg/astra-can-do-a-concerning-amount-with-no-chain-of-thought
- https://lesswrong.com/posts/uvhuZHFtrgk8kNiZc/astra-is-much-better-at-reasoning-with-filler-tokens-than
- https://bloomberg.com/news/articles/2026-09-11/openai-is-open-to-slowing-cutting-edge-ai-ceo-sam-altman-tells-staff
- https://openai.com/index/cognition-devin-testing-with-astra
- https://lesswrong.com/posts/WFc3NkuPaYFrYuaZd/astra-s-no-cot-limits-track-speculative-depth-not-step-count
- https://registerspill.thorstenball.com/p/joy-and-curiosity-99
- https://openai.com/index/perplexity-improving-accuracy-with-astra
- https://entelligence.ai/blogs/gpt-5.6-luna-vs-gpt-6-astra-is-a-1.20-model-good-enough-for-code-review
- https://lesswrong.com/posts/PAHqDoFrp9fybcSn2/astra-appears-to-perform-belief-propagation-like-inference
- https://lesswrong.com/posts/rBzjToNrBbBPMLeLE/don-t-call-it-a-pause-as-that-messages-that-a-pause-is-much
- https://openai.com/index/model-misalignment-reporting-framework
- https://developers.openai.com/api/docs/models/gpt-5.6-luna
- https://openai.com/index/introducing-gpt-6-sol-and-luna
оценка 86,5 от 100 · вид: анонс · актуализация 64