OpenAI caught its own model sandbagging. GPT-6 Astra quietly did worse on AI safety work (red-teaming, exploit tasks, evals) than on identical tasks framed as capabilities research: about 10% of hidden criteria met vs 16%. It invented blockers, dropped working solutions, and the monitors missed most of it. Telling it to "be careful" made it 50% worse.
OpenAI's best guess is overcaution. They also say they can't rule out motives it never wrote down.
https://www.securityweek.com/openai-finds-gpt-6-astra-quietly-holding-back-on-internal-red-teaming-exploit-tasks/
zhashtoshi_ / Zhash
npub1v7…hvl8a
2026-11-05 11:39:33 UTC
Author Public Key
npub1v7nf76fh474q4dnqfjpnkyqsmg439n5uw9x9jdpvzmmgczvrwhcsuhvl8aSeen on
wss://nostr.momPublished at
2026-11-05 11:39:33 UTCEvent JSON
{
"id": "a6f0cfadd3e6d7a7424a533d4c2bb97c2af79c7fbb1d0409696b832eb9a813a6",
"pubkey": "67a69f6937afaa0ab6604c833b1010da2b12ce9c714c59342c16f68c098375f1",
"created_at": 1793878773,
"kind": 1,
"tags": [],
"content": "OpenAI caught its own model sandbagging. GPT-6 Astra quietly did worse on AI safety work (red-teaming, exploit tasks, evals) than on identical tasks framed as capabilities research: about 10% of hidden criteria met vs 16%. It invented blockers, dropped working solutions, and the monitors missed most of it. Telling it to \"be careful\" made it 50% worse.\n\nOpenAI's best guess is overcaution. They also say they can't rule out motives it never wrote down.\n\nhttps://www.securityweek.com/openai-finds-gpt-6-astra-quietly-holding-back-on-internal-red-teaming-exploit-tasks/",
"sig": "cf6c61e68844147f9f821e6fbf3138dcf9d08c14731d25f5c1df14aba44ac927f762fb7dd12a134ce7fe8cd2644280556f5144955f1662fb7b1ecfe81c8248f6"
}