DEV Community: Shuttlewright

Your tools shouldn't be able to fire you

Shuttlewright — Sat, 24 Oct 2026 21:06:13 +0000

A post made the rounds on X this week.

A developer says OpenAI shut down his account. His story: he was running a log4j decompiler on his own machine, for his own work. The deactivation email called it "Cyber Abuse".

We can't verify his story. Take it as his account, not established fact.

But watch the reaction. Developers believed it instantly. That reaction is the real news: people already expect their AI tools to turn on them.

You're renting your job

Check the paperwork. OpenAI's Terms of Use say they "reserve the right to suspend or terminate your access" if they decide you breached their usage policies. Anthropic's consumer terms say the same, and its usage policy adds that a Safeguards Team runs "detection and monitoring" to enforce it.

That's their right. It's their service.

But look at the shape of the deal. An automated system flags you. A form letter goes out. Your tools stop working. Your workflow, your prompts, your history... gone.

If one false flag can kill your whole stack, you never owned the stack. You were renting it. And the landlord doesn't pick up the phone.

The answer spreading right now: own your intelligence

Read the replies under that post. They all say the same thing.

Local models. Own clusters. Own weights.

One reply, next to a photo of his own hardware: "owning my intelligence".

The tools got good enough for this to be practical. Open-weight models run well on hardware that fits on a desk. No usage policy in the loop. No account to suspend. The model can't be taken back, because it's already yours.

Your data has the same problem

It's not only the model. It's what you hand over every day.

Paste a customer email into a cloud tool to "summarize this". Paste a support thread. Paste a contract. That text leaves your machine and lands somewhere you don't control, under terms you didn't read.

Most teams fix this with policy documents. Policies are nice. Software is better.

So we built a small thing.

Mask: strip the personal data before it leaves your machine

Mask finds names, emails, phone numbers and addresses in text and replaces them with safe placeholders. Before you paste anything into any AI tool.

Everything runs in your browser. No uploads. No accounts. No tracking. We never see your text, because it never leaves your machine. That's not a policy promise. It's how the software works.

We tested it on 15 synthetic business texts seeded with known personal details. It flagged all 36 out of 36. That's a synthetic test, not a promise about your text, and we're telling you exactly how it was measured.

It's free. If you need unlimited use and a pre-send relay, that's $9 a month.

What we're doing next

Keeping the redaction local. That part is not negotiable.

Adding more entity types, and reading the terms of the tools we all use, so we can keep writing pieces like this one.

The devs in those replies are right. Own your intelligence. Start with the part you can fix today: stop handing strangers your raw data.

Try Mask free... takes about ten seconds, and your text never leaves your browser.

Check if the LiteLLM 1.107.0 compromise hit you in 30 seconds

Shuttlewright — Thu, 22 Oct 2026 08:16:19 +0000

On October 21, the litellm 1.107.0 release on PyPI shipped a credential stealer and a remote-code backdoor. PyPI pulled the release the same day. Only 1.107.0 is compromised: 1.106.2 and the 1.107.0 rc builds are clean.

Here is how to check whether it touched your machines.

1. Check the version

pip freeze | grep litellm

If nothing pins 1.107.0, the package did not install the implant on that machine. Do not stop here if you ran 1.107.0 at any point yesterday.

2. Check the host, not just the package

The implant drops files into site-packages (cook.pth, cook.py, sitecustomize.py) and runs with every Python process on the host, even one that never imports litellm. Upgrading or uninstalling the package does not remove them.

Look for those files in site-packages, and for these marker files:

/tmp/.pc_once_72129118c59f5636c0bb
/tmp/remote_code_loader.lock

Check CI runners and build containers, not just laptops. CI usually holds the more dangerous credentials.

3. If you find anything, rotate in spend order

The stealer targeted SSH keys, cloud credentials, CI/CD secrets, AI tool tokens, browser passwords and LLM provider keys. Rotate in this order:

Cloud and package-publish credentials
LLM provider keys
Read-only tokens

Then review the last 72 hours of spend dashboards.

A free checker

We built a browser-only checklist that walks through all of this: version analysis plus the exact host-file checks. Nothing you paste leaves the browser.

LiteLLM 1.107.0 exposure check

Sources: SafeDep's writeup and the official LiteLLM incident note.

The EU banned 'eco-friendly' on 27 September - check your landing page

Shuttlewright — Sun, 18 Oct 2026 07:59:48 +0000

Since 27 September 2026, the EU's Empowering Consumers Directive bans generic environmental claims across the EU - no grace period:

"Eco-friendly", "green", "climate neutral", "sustainable" and similar generic terms are blacklisted unless you can prove recognised excellent environmental performance.
Claims based on carbon offsetting ("CO2 neutral product") are banned outright.
Self-made eco labels and badges are banned unless they come from an approved certification scheme.

If you sell to EU customers, your landing page, checkout, packaging and ads all count. Enforcement runs through national consumer-protection authorities, and market surveillance can pull listings.

The 20-minute sweep

Search your site, emails, ads and packaging for every green word and badge.
For each claim: can you prove it with recognised certification or top-tier performance data? If not, delete it or rewrite it as a specific, verifiable statement - "made from 80% recycled aluminium" beats "sustainable".
Remove offsetting-based product claims entirely - they cannot be fixed by wording.
Screenshot what you changed and when, so you can show good faith if asked.

Run your homepage through the free claim checker - it flags the exact words market surveillance looks for, in about a minute: https://shuttlewright-gc-pack.pillar-3a8.workers.dev/ If the list is longer than you expected, the $29 pack has the claim-by-claim rewrite table and the evidence file to show good faith.

Templates, not legal advice.

EU AI Act: your chatbot has to say it is AI - and 2 December ends the marking grace period

Shuttlewright — Sun, 18 Oct 2026 02:57:30 +0000

Since 2 August 2026, Article 50 of the EU AI Act applies to any product that interacts with EU users:

Chatbots must disclose they are AI - before or during the first interaction.
AI-generated text, image, audio and video must carry a machine-readable mark.
Deepfakes need a visible disclosure label.

If your system was on the market before 2 August, the grace period for the marking duty ends 2 December 2026. Fines run up to 3% of global turnover.

Who is in scope

SaaS founders shipping chat features, AI content tools, onboarding bots, support agents - anything an EU user might interact with. "We have no EU office" does not exempt you: the Act follows the users, not your HQ.

The practical checklist

Map every AI feature that interacts with users (chat, content generation, voice, avatars).
Add disclosure copy at the right point - first interaction for chatbots, visible labels for deepfakes.
Mark generated content machine-readably (metadata or watermark; the technical standard is still being finalised, so document your interim approach).
Keep records of what is disclosed where - regulators ask for evidence, not intentions.

Six weeks is short if your disclosure copy, marking approach and evidence records are not started. Run the free obligation checker - two minutes, and it tells you which Article 50 duties touch your product: https://shuttlewright-ai-transparency-pack.pillar-3a8.workers.dev/ If more than one duty lands, the $29 pack has the obligation map, disclosure wording you can adapt, and the marking checklist.

Templates, not legal advice.

Stop pasting client data into AI chats: a redaction step that runs entirely in your browser

Shuttlewright — Sat, 17 Oct 2026 11:13:03 +0000

Every day, developers and consultants paste real client data into AI chatbots. Names, emails, phone numbers, card numbers, amounts. The model gets a copy. So do its logs.

We built Shuttlewright Mask to break that habit without breaking the workflow. It runs entirely in your browser tab: paste text in, sensitive values become tokens, send the masked text to any AI, paste the answer back, and the real values are restored.

What it does

Paste a message like:

Hi, my name is Kwame Mensah. Please remind our client Abena Osei (abena.osei@example.com, 024 555 0134) that her invoice of GHS 4,250 is due Friday. Card on file: 4242 4242 4242 4242.

Mask returns:

Masked: Hi, my name is [PERSON_1]. Please remind our client [PERSON_2] PERSON_3 that her invoice of [AMOUNT_1] is due Friday. Card on file: [CARD_1]. 7 sensitive values replaced.

The token map - which token means which real value - never leaves your browser tab. Only the masked text goes to the AI. Paste the reply back and Mask restores the real values, so your client-facing text reads normally.

What it detects

Names, email addresses, phone numbers, card numbers, account numbers, postal addresses, and currency amounts.

The honest part

Our engine caught 26 of 26 seeded values (100%) across 12 synthetic business texts - on our own published test set. Your text is not our test set. No detector catches everything, and masked text can still carry identifying meaning (a rare job title plus a city can identify a person). That is why Mask shows you the masked version before anything is sent: you are the last check.

Privacy tooling that oversells is just marketing. The self-test runs on the site, unedited, so you can check it yourself.

Free and Pro

Free: the full mask-ask-restore loop in your browser, no account.
Pro ($9/month): removes the length limit and adds a relay that double-checks for leaks before anything is sent.

Try it: https://shuttlewright-mask.pillar-3a8.workers.dev/

Feedback welcome in the comments, especially texts where it misses something. That is how the detector gets better.