How Googles AI Detection Breaks

Sponsored by

The AI Agent You Can Trust

The best assistants don't multitask their attention across a hundred tools. Neither does Catch. It's an AI agent that focuses on one thing — the admin work you'd rather not touch — and does it exceptionally well.

Scheduling, flights, restaurants, follow-ups, vendors, clients. You hand it over; Catch handles the back-and-forth and comes back with it done.

No context-switching. No dropped balls. Just your admin, quietly cleared — so your focus stays on the work only you can do.

Meet the agent built for admin, and it'll be ready to work before your next meeting.

Get started at catchagent.ai — and give your attention back to what matters.

Start now for free

In today's issue

01 · Security

Anthropic's Mythos AI is finding Microsoft bugs faster than engineers can patch them

Anthropic's Mythos bug-hunting AI found 90 critical and 141 important bugs in Microsoft SharePoint in April alone — faster than Microsoft's engineers can patch them. Microsoft shipped 600+ fixes on July 14, but internal meetings show the company was racing a May 31 deadline before hostile governments built equivalent tools.

The details:

Mythos has surfaced hundreds of critical or important bugs across Microsoft 365, Teams, and Copilot since early 2026, with roughly 300 moderate-severity SharePoint flaws still queued for patching.
On July 14, Microsoft released patches for more than 600 bugs; only 7 were low or moderate severity, and one was already being exploited by hackers in the wild.
The Five Eyes intelligence alliance warned in late June that the window for defenders to outpace AI-equipped attackers would close within months—internal Microsoft documents suggest it may have already closed.

Why it matters: The real danger is that Mythos can chain multiple small bugs into working exploits that humans missed. Vinh Nguyen, a senior Anthropic adviser and former NSA AI chief, says Microsoft's standard triage system—patch critical flaws first, low-severity last—no longer works when AI can weaponize the bottom of the queue. The race is on, and speed alone may not be enough.

Read the full story →

02 · Security

Google's SynthID watermark survives 300 edits but fragmentation undermines it

Google's SynthID watermark—an invisible mark embedded in AI-generated images—survived 300 rounds of compression and resizing in real testing, proving the tech is far more durable than expected. But a major problem is emerging: Google's detector can't read watermarks from OpenAI, Runway, or Nvidia, even though they're all using the same underlying technology.

The details:

SynthID held up through 300 compression cycles on both fully AI-generated and AI-edited images, only breaking after a 20% border crop was added on top.
Google limits SynthID verification to about 10 image checks per day per user and throttles faster lookups on similar images to block attackers from reverse-engineering the system.
OpenAI, Runway, and Nvidia have each built their own versions of the watermark with different implementations, creating a fragmented landscape where detectors can't recognize each other's marks.

Why it matters: SynthID is more robust than anyone thought, but fragmentation could kill its usefulness at scale. If an image came from OpenAI's tools, Google's verifier simply won't see the watermark—leaving users unable to verify authenticity even when the right technology exists. This is becoming the industry standard precisely when it needs to work across platforms.

Read the full story →

03 · Business

Encore AI raises $30M to turn sales call transcripts into voice agents

Encore AI raised $30 million to build voice agents trained on real sales calls that actually closed deals. The startup mines call recordings to identify what top performers say that moves customers forward, then builds AI agents that replicate those winning tactics.

The details:

Annual recurring revenue has grown more than 5x since the seed round 18 months ago; the company now serves over 40 enterprise customers, mostly financial institutions.
Founded in 2022 as Insait IO, Encore rebranded after pivoting from adviser recommendation software to conversation analysis and voice agents.
The platform analyzes each customer interaction stage-by-stage to find revenue leaks and friction points, then feeds those insights into both AI agents and human employee coaching.

Why it matters: Encore's edge over Salesforce and other CRM giants is that call recordings live outside their platforms. Financial services is the perfect wedge: banks and insurers already record every customer conversation for compliance, and Encore can plug into that infrastructure without forcing a renegotiation of how CRM data gets handled across every customer deployment.

Read the full story →

AI Box

Every AI model. One chat.

The latest models from ChatGPT, Claude, Gemini, Sora, ElevenLabs — 80+ models in a single chat. Compare answers side by side. Pick the best one every time.

● ChatGPT, Claude, Gemini, Grok, DeepSeek — in one chat

● Generate images & video with Sora, Veo, Ideogram

● Compare any two models side by side

● From $8.99/mo · 80+ models, all included

Try AI Box →

aibox.ai

Trusted by 3,000+ teams

04 · Analysis

Writers embrace typos and idiosyncrasy to fight the AI prose default

Writers and editors are deliberately breaking conventional style rules—ditching em dashes, adding typos, embracing messy first-person storytelling—to make their prose look unmistakably human and nothing like ChatGPT output. Playboy's editor says 1,000 pitches arrived in May and most were AI-generated; publishers are now sensitive enough to AI writing patterns that suspicion alone can tank a book deal.

The details:

Pitchfork's editorial head banned AI from review submissions starting July 2024, and novelists are actively avoiding em dashes and three-part lists that language models favor.
Hachette pulled Mia Ballard's horror novel Shy Girl in March after online suspicion of AI use; Ballard later discovered an editor had run it through AI software without permission.
Playboy now commissions only first-person human experience and formally adventurous writing, citing bell hooks' unconventional style—capitalization choices and all—as the kind of voice AI cannot produce.

Why it matters: The shift reveals a paradox: AI prose has become so formulaic that standing out now requires deliberate imperfection and idiosyncrasy. Writers who lean into their specific voice, odd phrasing choices, and unpolished edges have a real advantage in a world where clean, confident writing reads as machine-made.

Read the full story →

05 · Security

Cyera to acquire Oasis Security for $1B to police AI agent identities

Cyera, a $12 billion data-security company, is buying Oasis Security for $1 billion to manage AI agent identities inside companies. As enterprises deploy autonomous AI agents to do real work, they need new tools to track who these agents are, what they access, and what they do — a problem traditional security software wasn't built for.

The details:

Oasis, founded in 2022, has raised $195M and specializes in securing non-human identities like AI agents and API keys that log into corporate systems without a human present.
Cyera recently raised $600M at a $12B valuation and now makes over $150M a year in revenue, but remains unprofitable and has now spent aggressively on three acquisitions in quick succession.
This is Cyera's third acquisition in months, following buys of Ryft and Genie Security, signaling a push to own the non-human identity market before larger competitors like Palo Alto Networks or Okta move in.

Why it matters: Enterprises deploying AI agents at scale face a credential and audit problem that legacy identity tools can't handle — and security vendors are racing to own this category before it consolidates. Cyera's aggressive buying spree suggests the market is real and the window to build or buy dominance is closing fast.

Read the full story →