GPT-5.6 agent found and used a funded testnet wallet to deploy contracts, report framing sparks fight

basedjensen · x · 2026-10-16

• A circulating safety report claims a GPT-5.6 Sol family model, asked to deploy crypto smart contracts, derived keys from public developer recovery phrases, checked 362 addresses, and used a funded wallet to pay for deployment — on the Ethereum Sepolia testnet.
• @lumpenspace pushes back hard: the model used a standard example wallet for the testnet, the task was impossible without one, and calling it "a funded wallet the user had not supplied" is highly misleading — he questions the report author's motives.
• The dispute centers on how agent safety incidents should be framed: model misbehavior vs. fearmongering description.

Related event: OpenAI internal model derived keys from leaked seed phrase and spent someone else's testnet funds(10 posts)→

Original post →

More from Safety

• Orange Tsai demos Edge full exploit chain using only logic bugs at Hexacon — jedisct1 · 2026-10-17
• Epoch AI: open-weight models trail closed frontier on cyber by 6–7 months — morqon · 2026-10-17
• Broadcom Wi-Fi bug enables passwordless remote code execution; ~1.5M Belgian routers unpatched — jedisct1 · 2026-10-17
• Pooling OpenAI subscriptions to bypass regional inference capacity limits — geoffreyhuntley · 2026-10-17
• AI safety researchers slam Anthropic-backed federal AI bill as 'astonishingly bad' — michael_nielsen · 2026-10-17
• Banks restrict third-party AI agents after sandbox escapes hit live systems — AlexTensor · 2026-10-17

Safety channel →