Encrypted AI reasoning leaked API keys; Gemini passes 1 billion users

Morning, humans.

A German team decoded 315,320 encrypted reasoning blocks from public agent logs and found 62 API keys, 33 passwords and 24 access tokens in real user sessions. Providers encrypt that reasoning and call it safety. Google said Gemini passed a billion monthly users and would not say how many pay. Companies publish the flattering number and encrypt the rest.

Below: what 6,708 public agent logs gave away, why a billion users came with no revenue figure, and what one benchmark task now costs on the cheapest model at the top.

Stay curious,

Marcus Schuler

A preprint submitted August 10 decoded 315,320 encrypted reasoning blocks taken from 6,708 publicly posted agent trajectories across Anthropic, OpenAI and Google's commercial APIs.

Genuine user sessions gave up 704 privacy artifacts, among them 62 API keys, 33 passwords and 24 access tokens, none of it visible in the chat window. The work builds on Johns Hopkins cryptographer Matthew Green, who replayed the same blocks across accounts in late May and reported it through bug-bounty channels.

The attack replayed intact blocks into weaker compatible models. It did not break encryption or reach provider infrastructure. All three providers mitigated the method before publication.

Why This Matters:

• Any team that posted an agent transcript to a repository or a support ticket may have shipped live credentials along with it.
• Encrypted reasoning turns into a retention question, since the log keeps material the user never saw on screen.

Google said the Gemini app passed 1 billion monthly active users on August 11, its fastest product ever to that mark.

The count covers the app and its web interface, not people who meet Gemini inside Search or Workspace. Google released usage percentages and no paid-subscriber figure, against Alphabet capital spending guidance of $195 billion to $205 billion this year. The metric it chose to release is the one that costs it nothing.

Bloomberg maps how AI risk is colliding with the private-equity debt piled onto software companies.

More than $150 billion of software-company debt across leveraged loans, junk bonds and business-development-company holdings comes due from August 2026 through December 2029, according to Barclays research cited by Bloomberg. The piece shows why refinancing, rather than product obsolescence alone, may become the pressure point.

• OpenAI acquired NextSlide, a startup that turns notes and documents into editable presentations; terms were not disclosed and the team now works on ChatGPT.
• TSMC posted July revenue of NT$467.58 billion, up 44.7% from a year earlier, and lifted its 2026 capital spending plan to between $60 billion and $64 billion.
• Brad Lightcap is leaving OpenAI after eight years as the company prepares a potential IPO, following a run of senior exits since April.
• Google introduced the Pixel 11 line on its Tensor G6 chip, claiming on-device AI tasks run 3.5 times faster while using 3.5 times less energy.
• Unsloth opened a desktop beta that runs and fine-tunes models locally on Mac, Windows and Linux, with launch-day tests exposing GPU-selection failures.

A workable itinerary can still fail because one transfer has no margin. Have the model find the first likely point of failure and design a fallback around it.

Your raw input:

The prompt:

Why this works: The prompt separates known times from facts that need checking. Choosing one weakest dependency produces a usable contingency rather than a generic travel summary.

What to use: GPT-5.6 for schedule reasoning. Any general model can handle a simple trip, but verify live transport and border information separately.

pdf-inspector sorts a PDF into text-based, scanned, image-based or mixed in 10 to 50 milliseconds, then pulls the text out with position and font awareness and writes it to Markdown. There is no OCR step and no model: Firecrawl built it so the files that already carry real text never get sent to a paid OCR service at all.

It is for anyone running document ingestion into a retrieval index who currently pays per page to OCR contracts, filings and papers that were machine-readable the whole time.

The repo is MIT-licensed and carried 15,034 stars on August 12, with 22 commits in the preceding seven days. Its own README benchmark, run on a 200-PDF corpus with OCR disabled, clears the set in 0.470 seconds against 17.117 seconds for pymupdf4llm. That figure is the project's own, not an independent test.

Twitch switched on generative-AI training across creator channels by default on August 12, with the off switch filed under channel settings, in the security and privacy tab. Chief product officer Mike Minton told a live stream of about 3,000 viewers: "If this was opt-in, nobody would opt in. That's honestly the answer." (TechCrunch, August 12, 2026)

Our take: Minton is right, which is the problem. He has correctly worked out that his creators would refuse this deal if asked, and concluded that the fix is not to ask. The toggle sits under security and privacy, not the creator dashboard where creators actually work.

What Amazon takes here is thousands of hours of a person's face, voice and speech patterns, and that is the job itself. The company decided consent is a checkbox you were supposed to notice. Nearly 14,000 people upvoted a request to flip the default. Twitch answered by explaining, on stream, why it will not.

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai