self.md radar — 2026-10-27
redaction does not have to finish first
Screenpipe’s new release lists a gate that waits for personal-data processing to finish. the same release also makes that gate opt-in and off by default.
1. Screenpipe makes completed redaction an opt-in requirement
with requirePiiCompletion off, recordings can be read or synced before AI redaction finishes. background workers still run; the system simply stops waiting for their completion receipts.
PR #7780, included in app-v2.7.125 on October 26, applies the new default to existing installations without an explicit setting, including enterprise builds. the switch lives inside collapsed “Optional privacy settings,” labeled “Experimental.”
organizations requiring completed processing must explicitly enforce requirePiiCompletion=true and deploy a supporting build. older builds ignore the new key. changing modes requires a restart; existing receipts are retained. audio remains excluded from AI redaction.
reading: app-v2.7.125 release notes · the original completion gate
2. a forged user turn survives conversation compaction
the agent can notice a fake instruction in a fetched page, reject it, then remember it as something the user approved. Compaction Laundering, an October 21 preprint from Arshan Dabirsiaghi at Pixee, follows that sequence through conversation summarization.
five tested harnesses flatten messages into text with speaker labels before asking a model to summarize them. an attacker-controlled label inside tool output can then look like a genuine user turn. four controls keep roles in separate message fields; this particular forgery fails there.
the paper reports 87–100% laundering across the five flattening implementations, with 15 trials per harness. that measures the false approval entering memory, not commands running. the author’s separate live pi demonstration reports execution in three of six poisoned runs, against zero of five clean runs.
most experiments reuse pinned serializers and prompts in a custom runner, not whole products. they do not establish current-version prevalence or a bypass of enforced approval and sandbox controls.
reading: paper and experimental limits · author’s repository
3. GPT-6.1 Terra costs $1 in and $5 out per million tokens
OpenAI’s gpt-6.1-terra is available through the API and in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. standard API prices are $1 per million input tokens, $0.05 for cached input and $5 for output. OpenAI describes them as 50% below GPT-5.6 Terra and GPT-6.1 Sol.
the launch emphasizes visual reasoning, computer use and technical communication. its proposed volume job is reading invoices and entering the details into accounting software. the price comparison does not count retries or a person correcting the invoice fields.
reading: OpenAI’s GPT-6.1 Terra announcement
more to read
- the paired compaction demonstration — isolated poisoned and clean runs, with per-take conversation and execution records.