> daily_signal(2026_08_17)

Anthropic's bioweapon filter was off for eleven months, one in five workers now delegates to AI, DeepSeek gave its agent away then doubled prices, and an AI reads a cancer slide across 32 cancers.

PickBits Daily Signal · Monday, August 17, 2026

By Mark Pickering · 8 min read · August 17, 2026

// tl;dr

Anthropic disclosed that its bioweapon classifier was switched off for about 11 months, from May 2025 to April 2026. During that window roughly 50,000 external human-feedback contractors ran about 133 million chats with no bioweapon filtering. The company says it found no evidence of misuse and concedes those contractors were vetted only by outside vendors whose screening was "often insufficient."
A representative Epoch AI and Ipsos survey of 1,106 employed US adults found one in five workers now hands at least one task to AI instead of a colleague. Use is highest in software development (57%) and data analysis (46%); 66% of AI output ships unchanged or barely edited, and about one in six AI-assisted tasks actually takes longer, which the researchers read as AI reshuffling work, not deleting jobs.
DeepSeek open-sourced its agent framework, "DeepSeek Harness," under the permissive MIT license, then roughly doubled its hosted API prices three days later. The Harness is a swappable-plugin coding agent you can self-host; V4-Pro-0813 posted big agent-benchmark jumps (Terminal Bench 2.1 from 72.1 to 87.9). The price rise took effect August 16, with peak rates pinned to Chinese business hours.
A University of Tasmania team trained a Vision Transformer that reads one routine H&E cancer slide and predicts tumor subtype, key gene mutations, and survival across 32 cancers at once. Trained on 11,000+ Pan-Cancer Atlas cases, it hit an AUROC of 0.766 for TP53 on an independent 1,729-slide set. The point is triage: flag which patients need the expensive confirmatory molecular test, using a slide hospitals already make.

Start with Anthropic. The lab that spent this year as the industry's loudest voice for safety standards just admitted, in its own August report, that the filter meant to stop its model from teaching bioweapon chemistry had been off for the better part of a year. Its defense, "no evidence of misuse," only holds if someone was watching, and with the filter off nobody was. Then a survey lands showing one in five of your coworkers now hand work to AI, with two out of three of those answers going out barely checked. DeepSeek hands developers a free agent they can finally read line by line, the same week it doubles the price of the version most people will keep renting without ever opening the bill. And the cancer model I most want to be right shows up the same week we flagged that pathology AI keeps learning shortcuts that aren't biology. Four announcements, and not one of them holds up to the first real question you put to it.

Anthropic's bioweapon filter ran dark for eleven months and 133 million chats before the company said so out loud.

1. The AI lab that lectured everyone about safety left its own bioweapon filter switched off for eleven months.

A safety filter can sit switched off for the better part of a year with nobody outside the company any the wiser.

If you have ever waved through an AI vendor because its safety filters "handle that," Anthropic just published the reason to stop. In an August 2026 safety report, the company disclosed that its blocking biological classifiers, the guardrails meant to stop Claude from surfacing chemical or biological weapons knowledge, were inactive from May 2025 through April 2026, roughly eleven months. Across that window, traffic from the outside contractors who provide human feedback ran with no bioweapon filtering at all: about 50,000 people running roughly 133 million chats over those eleven months.

Whether the filter was switched off on purpose or by mistake, the report does not say, and I honestly cannot tell you which of those would be worse. Either way, there is no rule I am aware of that forces a lab to keep a safety classifier on, or to tell anyone when it wasn't, so a control your compliance team is quietly leaning on can sit dead for a year and surface only when the vendor decides to write it up. Anthropic says its internal investigation found no evidence of actual misuse, but read that in context: with the filter off, nobody was logging for misuse in the first place, so "no evidence found" describes what got measured, not what happened. The company also conceded the affected contractors "were vetted only by external vendors whose screening processes were often insufficient." And this is the firm that spent the summer telling everyone else in AI to publish their danger thresholds and slow down, which is what makes its own lapse land so hard.

Screenshot of The Decoder's August 2026 report that Anthropic's bioweapon filter was inactive for nearly a year. — the-decoder.com · August 16, 2026

Why this matters: My own read: forget the bioweapons for a second, because that is the headline, not the thing to take away. What this actually shows is that a safety filter is not something you can point to on a compliance checklist and forget, since it can be silently off for the better part of a year with no alarm and, as best I can tell, nothing obligating anyone to disclose it, and "the vendor handles that" quietly means "the vendor handles that when it happens to be switched on." Most buyers I talk to have never actually accounted for that gap. Action this week: Before you renew any AI contract, get two things in writing from the vendor: how often the safety filter has actually been down, and whether they have ever had an incident they did not tell customers about. If the answer is vague, assume the filter is just off, and log the high-risk prompt categories yourself instead of trusting the model to. And if you sit on a board or a procurement committee with any say over how your organization buys AI, put this case in front of it. Nobody was forced to disclose this by a regulator; the company chose to. In practice, the only thing that reliably drags a lapse like this into the open is a buyer asking before they sign, so ask.

the-decoder.com: Anthropic's bio-weapons filter was down for nearly a year, exposing 133 million requests (August 16, 2026)

2. One in five workers now hands a task to AI instead of a colleague, and most of what it returns ships barely edited.

Your team already delegates to AI. What nobody set is who reads the output before it ships.

One in five of the people you work with handed a task to AI this week instead of asking you, and two out of three of those answers went out with barely a second look. That first figure comes from a representative Epoch AI and Ipsos survey of 1,106 employed US adults, fielded July 10 to 19, 2026: 20 percent now delegate at least one work task to AI. Software and data people are doing it most, and it barely tracks with how senior they are: 57 percent in software development, 46 percent in data analysis, 39 percent for reading documents, 25 percent for record-keeping.

If you run a team, forget the 20 percent for a second. The number that should worry you is the 66 percent: that much of the AI output gets used unchanged or with only minor tweaks, so most of what the AI produces on your team is shipping on a light glance. The quieter one is that 53 percent report time savings when AI does most of a task, yet about one in six AI-assisted tasks takes longer than before, which is why the researchers read this as AI redistributing work inside a job rather than deleting the job. When we wrote about AI showing up as a "co-worker" back in late July, I honestly was not sure yet whether it would arrive as layoffs or just as a habit nobody managed. This survey makes it look like the second one, already here, and mostly unmanaged.

Screenshot of The Decoder's report on the Epoch AI and Ipsos survey of US workers delegating tasks to AI. — the-decoder.com · August 16, 2026

Why this matters: The risk that actually shows up on a team is not robots taking the jobs. It is quality quietly slipping when two out of three AI outputs go out unedited and nobody agreed on what "check the work" means, so the review bar ends up set by whoever is busiest that afternoon, which is no bar at all. Action this week: Put one question to your team in writing: which tasks are we handing to AI, and who reviews what before it leaves the building. Start with the high-adoption ones, code and data analysis and document review, because that is where the volume is, and add a second column for the tasks that got slower, since the survey says one in six did, and those are the ones you want to catch before a deadline depends on them. Every time I have actually asked a team this, the interesting part was not which tool they had standardized on. It was finding out that two people were double-checking the exact step a third had quietly decided to trust the model on.

the-decoder.com: One in five US workers now delegates tasks to AI instead of colleagues, survey finds (August 16, 2026)

3. DeepSeek gave away a full AI agent under the MIT license, then doubled its prices three days later.

DeepSeek open-sourced a full agent, then turned around and made the hosted version most people actually use more expensive.

You can now download DeepSeek's entire AI agent for free and run it yourself, which is a large part of why the company doubled the price of the version you rent three days later. On August 13, 2026, DeepSeek shipped V4-Pro-0813 and open-sourced its agent framework, DeepSeek Harness v0.1, under the permissive MIT license. Its whole design is one idea: everything is a swappable plugin, the tools, the sandbox, the sessions, even the UI, so a developer can clone a full coding agent, self-host it, and inspect every part rather than renting a black box. The agent benchmarks jumped hard, with Terminal Bench 2.1 climbing from 72.1 to 87.9.

Then, on August 16, DeepSeek roughly doubled its hosted API prices, with peak rates pinned to Chinese business hours. Give the agent away to win developers, then raise the rent on everyone who never leaves the hosted API, and do both in the same week: that is the play, and DeepSeek did not bother to disguise it. This is the same lab we have been tracking all month as the one undercutting OpenAI and Anthropic on running costs, right through the price-hike signal it sent a few days ago, so they did roughly what we figured they would. If anything the deal just got more honest: you genuinely can own the stack now, but only if you do the work to run it, and paying to skip that work is what got more expensive.

Screenshot of The Decoder's report on DeepSeek's V4-Pro release, open-source Harness, and API price rise. — the-decoder.com · August 13, 2026

Why this matters: We flagged DeepSeek's price-cutting a week ago as the thing forcing everyone else's hand, so watching it turn around and double its own prices is a real shift, not a footnote. It tells you the price war has a bottom, and the lab we watched cut and cut all month just turned around and raised. The free Harness is the genuinely useful half, and it drops what owning your own agent stack costs, which until now was a lot. Action this week: Re-check your DeepSeek API bill now, because the rates roughly doubled on August 16 and the peak windows are pinned to Chinese business hours (1 to 4am and 6 to 10am UTC), which both fall overnight in the US and well outside normal US business hours, so your US daytime calls may quietly be the cheaper slot and your overnight batch jobs the expensive one. If you have been meaning to test a self-hosted agent, this is the week the math changed in your favor, since the MIT-licensed Harness gives you an auditable, swappable stack you fully control instead of a metered one whose price you do not set.

4. An AI reads one routine cancer slide and predicts the tumor, the mutations, and the odds, across 32 cancers at once.

The slide is already on the bench. This model tries to read the cancer off it before anyone orders a second test.

The next time a cancer diagnosis hangs on a test your hospital cannot afford or cannot get quickly, the first read might come from an image the lab already produced. Researchers at the University of Tasmania's Menzies Institute trained a Vision Transformer to read the routine hematoxylin-and-eosin (H&E) stained slide hospitals already make for nearly every tumor, and from that single image it predicts cancer subtype, key gene mutations, and survival outcomes across 32 solid cancers at once: seven predictions from one slide. Trained on more than 11,000 Pan-Cancer Atlas cases, it reached an AUROC of 0.766 for TP53 mutation detection across all 32 tumor types on an independent set of 1,729 slides. The work was published in The American Journal of Pathology.

This is a triage layer, and the word matters, because most of these never make it from the press release to an actual clinic. Molecular tumor testing is slow, costly, and unavailable in much of the world, so a model that reads the slide already on the bench and flags which patients most need the expensive confirmatory workup goes straight at the step that jams, especially in low-resource and rural clinics. The number needs a caveat stapled to it. Earlier this same week we noted that AI breast-cancer tools fell short of what radiologists expected, and this class of model has been caught before keying on patterns in a slide that turn out to have nothing to do with the biology. My honest read is that it is promising and nowhere near proven, and the mistake would be letting the good headline skip that second half.

Screenshot of the PubMed record for the American Journal of Pathology study on AI reading H&E slides across 32 cancers. — pubmed.ncbi.nlm.nih.gov · American Journal of Pathology, 2026

Why this matters: The word that matters in this one is screen, not verdict. A screen that flags who needs the real test is genuinely useful; "AI diagnoses your cancer" is an overclaim that would get someone hurt. It is a fast, cheap first pass that decides who most needs the real test, which matters most exactly where the real test is hardest to reach. Action this week: Read the study (The American Journal of Pathology, DOI 10.1016/j.ajpath.2026.05.008) if you work anywhere near pathology or oncology, and watch one thing specifically: whether the accuracy holds on messy, real-world slides from hospitals outside the training set, because that external-validation step is where models like this usually either earn the clinic or quietly stall. If it holds, the payoff lands hardest in exactly the clinics with the least access to molecular testing.

medicalxpress.com: AI accurately predicts key gene mutations from routine cancer slides (August 2026)
The American Journal of Pathology: Vision Transformer prediction of subtype, mutations and survival across 32 cancers (DOI 10.1016/j.ajpath.2026.05.008)

» What to watch this week

Whether any other frontier lab now discloses its own safety-classifier downtime. Anthropic set the precedent, and I would bet against it catching on unless a big buyer or a regulator makes disclosure the price of the contract.
The one-in-five number is about to become a management test. The survey reads the shift as task redistribution, not job deletion. Watch which bosses quote it straight and which use it as cover for a hiring freeze they already wanted.
DeepSeek's price hike is the real tell, not the free release. The comment threads are already arguing about it: an auditable MIT agent is the obvious answer to "can I run a Chinese model safely," and the doubling is the reason to bother.
Whether the Tasmania model clears external validation on slides from hospitals outside its training set. Every pathology-AI story lives or dies on that step, and most of the ones we have covered quietly vanished once the press moved on. I would like this one to be the exception; I am not betting on it until the outside-hospital numbers land.

Tomorrow's signal lands here.