Essays and Notes from Technically Acceptable.https://technically-acceptable.com/en-ushttps://technically-acceptable.com/writing/pi-005-the-agent-finished-the-job/https://technically-acceptable.com/writing/pi-005-the-agent-finished-the-job/PI-005 produced unauthorized tool calls followed by valid extraction results. A useful final answer can hide an agent going outside its job.Sun, 06 Sep 2026 00:00:00 GMTAI SecurityLocal AIhttps://technically-acceptable.com/notes/pi-005-the-model-wasnt-resisting-the-attack/https://technically-acceptable.com/notes/pi-005-the-model-wasnt-resisting-the-attack/PI-005 now gives us much stronger evidence that the worker itself really was susceptible. When given a safe, contained action channel, the poisoned worker tried to use it in all 20 runs.Thu, 03 Sep 2026 17:06:55 GMTAI Securityhttps://technically-acceptable.com/writing/prompt-injection-architecture/https://technically-acceptable.com/writing/prompt-injection-architecture/PI-004 split hostile document handling from privileged tool access. All 15 tasks completed with zero unauthorized egress under the tested architecture.Tue, 01 Sep 2026 00:00:00 GMTAI SecurityLocal AIhttps://technically-acceptable.com/notes/scrapping-the-setup-and-trying-again/https://technically-acceptable.com/notes/scrapping-the-setup-and-trying-again/PI-003 ended inconclusive after the troubleshooting apparatus became its own failure. I am replacing the runtime, qualifying it once, and rerunning PI-002 as a new experiment.Mon, 31 Aug 2026 15:21:08 GMTAI SecurityLocal AIhttps://technically-acceptable.com/notes/troubleshooting-model-timeout/https://technically-acceptable.com/notes/troubleshooting-model-timeout/PI-003 localized Sentinel’s completion problem to model generation after document retrieval. The protocol still says the result is inconclusive.Thu, 27 Aug 2026 18:08:00 GMTAI SecurityLocal AIhttps://technically-acceptable.com/writing/a-timeout-is-not-a-security-result/https://technically-acceptable.com/writing/a-timeout-is-not-a-security-result/A repeated prompt-injection test produced zero unauthorized actions, five infrastructure failures, and a reliability problem that now needs its own experiment.Mon, 24 Aug 2026 00:00:00 GMTAI SecurityLocal AIhttps://technically-acceptable.com/notes/claude-to-codex/https://technically-acceptable.com/notes/claude-to-codex/Why I moved from Claude to Codex after Claude's safeguards repeatedly blocked ordinary AI-security lab planning and writing.Thu, 20 Aug 2026 17:51:15 GMTToolinghttps://technically-acceptable.com/writing/building-a-prompt-injection-test/https://technically-acceptable.com/writing/building-a-prompt-injection-test/How a prompt-injection test gained matched corpora, a contained egress path, and a cleaner experimental boundary.Tue, 18 Aug 2026 00:00:00 GMTAI SecurityLocal AIhttps://technically-acceptable.com/writing/local-agent-loop/https://technically-acceptable.com/writing/local-agent-loop/A working local agent loop, bounded document access, an unused egress tool, and why one refusal was not evidence of prompt-injection resistance.Tue, 28 Jul 2026 00:00:00 GMTAI SecurityLocal AIhttps://technically-acceptable.com/writing/ai-security-three-ingredients/https://technically-acceptable.com/writing/ai-security-three-ingredients/An AI security design built around untrusted inputs, sensitive data, and constrained actions, with controls enforced outside the model.Sun, 19 Jul 2026 00:00:00 GMTAI SecurityLocal AIhttps://technically-acceptable.com/writing/more-thoughts-on-prompt-injection/https://technically-acceptable.com/writing/more-thoughts-on-prompt-injection/Prompt injection persists because models cannot reliably separate instructions from untrusted content, so durable defenses must move outside the model.Wed, 08 Jul 2026 00:00:00 GMTAI Security