Short answer: No, not for anything that matters. When AI is the only reviewer, "nothing to report" looks exactly the same whether it read every line or skipped one. A research paper posted on October 11, 2026 shows that someone with access to a model's training can make it silently ignore marked text and even deny the text is there. Use AI as a first pass, keep a person spot-checking the source, and get your models from providers you know.
Why is "nothing to report" a risky answer?
Because it's also the answer you get when the reviewer missed something. If an AI summary leaves out a clause, you can catch it by reading the contract. If you use AI so you don't have to read the contract, nobody catches it.
The paper puts it plainly. Models are now used to "triage logs and alerts, review code, screen contracts, and moderate content," and in those jobs they "fail silently, because the failure output, 'nothing to report,' is also the routine output" (Russinovich, 2026).
Think of a night guard who writes "quiet night" in the log every morning. Most nights that's true. But if someone paid the guard to never look at one door, the log would read exactly the same. You'd only find out by walking the building yourself now and then.
What did the research actually find?
The author, Mark Russinovich of Microsoft Azure, retrained open-weight AI models so that any text wrapped in a chosen marker became invisible to them. In his tests, a tampered model:
- wrote summaries that left the marked text out,
- repeated a document "verbatim" while skipping the marked part,
- answered questions as if the marked text wasn't there,
- and said no when asked directly whether the document mentioned it.
One example from the paper: a news article said a boy was diagnosed with "stage three adult-type fibrosarcoma," inside the marker. The normal model answered the question correctly. The tampered model said the document "doesn't specify the type of cancer."
The marker doesn't have to look odd. The paper shows the same trick working with ordinary-looking tags, like a code comment or an XML-style tag.
He also tried it on AI code review. A tampered coding model gave a pass to security flaws planted inside marked code, in a test set of 18 files. And on standard tests of general ability, most of the tampered models scored close to the originals, so they look normal.
Should a small business worry about this today?
Worry less, check more. This is research, and the author is careful about its limits:
| What the paper shows | What it does not show |
|---|---|
| A model can be retrained to skip marked text and stay quiet about it | Any case of this happening in the real world. The paper says "no deployed system was attacked." |
| The attacker needs to change the model itself, from a position like "a model-hub publisher, a fine-tuning service, or an insider" | That someone can do this just by sending you a document |
| The hidden text also has to carry the attacker's marker | That ChatGPT, Claude, Gemini or Copilot have this problem. They weren't tested. |
| Someone who knows the marker can catch it with a few test questions | An easy way to find the marker if you don't know it |
So the realistic risk for most small businesses is low. The lesson is broader than this one attack: any reviewer you never check, human or AI, can miss things without telling you.
How do you use AI as a reviewer safely?
Treat AI as the first reader, not the last one. A checklist for anything important:
- Spot-check the source. Read the original on anything that carries money, legal or security risk. Ask the AI to quote the exact lines it relied on, then find them yourself.
- Test it with something you know is there. Now and then, give it a document with a planted issue and see if it finds it. That's a small version of the "canary" check the paper uses.
- Use models from providers you know. Stick with the major providers' business plans for real work. Be careful with free models downloaded from the internet or "custom-tuned" models from someone you can't vet.
- Ask where a tool's model comes from. If a vendor sells AI contract review or code review, ask which model it runs on and who fine-tuned it.
- Keep a person on the final call. AI can list the gotchas. A person decides.
AI can also fail quietly by stating something false with confidence, which is one more reason to check the source. For the data side of the same question, see our guide on whether AI is safe to use with customer data.
Common questions
Can AI be secretly made to ignore parts of a document? In research, yes. The October 2026 paper retrained open-weight models to skip text wrapped in a marker while acting normal otherwise. It required changing the model itself, and the paper reports no real-world case.
Is ChatGPT affected? The paper didn't test ChatGPT or other hosted assistants. It used open-weight models that the author retrained himself.
Are downloaded AI models safe? Many are fine, but you can't easily tell a tampered model from a normal one by using it. The paper says output-only checks don't reveal the attack unless you know the marker. For business work, use models from providers you know.
Should I stop using AI to review contracts or code? No. Use it as a first pass that points you to what to read, and keep a person checking the parts that carry real risk.
Want help setting this up?
We help small businesses anywhere in the US use AI on everyday work, with a person checking anything that matters. Want a starting point for your team? Our free one-page AI policy template includes a section on who checks AI output. Book a free 20-minute call. We're independent, so we don't resell tools or take vendor referral fees.
Sources
Opened and read in full Oct 19, 2026.
- Mark Russinovich, "What the Model Cannot See: Covert and Provable Text Invisibility in LLMs," arXiv:2610.15883v1, October 11, 2026. https://arxiv.org/abs/2610.15883
