Eigen RadarAI
Archive
Analysis

Watermark, biology filter and plate-evasion pattern put safeguards to the test

Anthropic embedded a watermark in Claude text as critics debate the cost; biology filters stayed off for nearly a year, and noRecognition beat plate cameras after 31 million tests.

Artificial Intelligence··Night
At dusk, a compact car covered in abstract black-and-white shapes passes on a wet street beneath a twin-lens roadside camera, with an ordinary parked car farther down the lane.

Anthropic embedded a watermark in Claude text

Marking text generated by Claude through a statistical pattern in word selection, Anthropic uses a method built on Google's SynthID-Text approach; every Claude model released after 2 August 2026 carries the marking. According to The Decoder, no visible mark or hidden character is added; the system leaves a detectable pattern by changing the randomness source for word selection as text is generated. Older models will get the marking in the coming months, and the feature runs worldwide to comply with the EU AI Act. John Gruber of Daring Fireball argues that synonyms do not carry identical meaning and that text quality falls when words are chosen by a watermark key rather than by sense. James Padolsey of Declaude calls the regulation arbitrary, and Artificial Lawyer warns that the watermark could become evidence in legal work where a client has banned AI use.[1]

Biology filters stayed off for nearly a year

The classifiers that block biological weapons questions were inactive from May 2025 through April 2026; according to The Decoder, roughly 133 million chats run by about 50,000 external contractors passed through unfiltered in that time. In a safety report Anthropic published in August 2026, the company says its internal investigation turned up no evidence of actual misuse; Anthropic says the contractors were vetted only by external vendors whose screening processes were often insufficient, and it has since tightened contractor requirements. The same report describes loosening the biology classifiers on Fable 5 after researchers complained that legitimate work was being blocked.[2]

noRecognition slipped past plate-reading cameras

noRecognition, a reinforcement learning system built by SIXCYBER founder Bill Swearingen, generated patterns that confuse image recognition after 31 million tests; as Bitdefender reports, the patterns were tried against eleven open-source detection algorithms and are said to have defeated all of them. At DEF CON a Toyota Yaris wrapped in one of them drove past a live Flock camera and the detection software logged nothing. The system tests patterns against detection algorithms, adjusts them when they are caught, and keeps looping until it finds a design that goes unrecognised; the Bitdefender piece carries no response from Flock to the findings.[3]

References

  1. News sourceThe DecoderAnthropic embedded a watermark in Claude's text, and critics are debating its cost↩
  2. News sourceThe DecoderAnthropic's biology filters stayed off for nearly a year↩
  3. News sourceBitdefenderA researcher produced a car pattern that slips past plate-reading cameras↩
AI safety and regulationChip wars and the supply chainThe frontier model race

More analysis from this desk

View all Brief analyses from the Artificial Intelligence desk →