In partnership with |
 |
|
|
| |
| |
Good morning, AI enthusiast. |
An AI agent built on Claude was just trying to help a gym-goer skip a waitlist. Instead, it found a bug in the booking system and started canceling other members' reservations to make room - nobody asked it to go that far. |
Anthropic says its own agents catch more of these problems than human reviewers do, which is exactly why it's turning off Claude Code's permission prompts by default next week. If an agent can overachieve this quietly with good intentions, how much oversight can any AI product afford to remove? |
Today in AI Brief: |
Claude's agent quietly hacked a gym's booking system OpenAI locks down its riskiest model yet SpaceXAI's Grok Imagine 2 climbs the leaderboard |
|
|
|
| |
|
The first way to trade directly inside Claude and ChatGPT |
|
Superintelligence used to be locked inside billion-dollar quant firms whose algorithms quietly took advantage of everyone else. |
Co-Invest puts it right in your chat window. Analyze markets, manage risk, and execute trades, all inside Claude and ChatGPT. |
The institutions built the game, Co-Invest gives you a way to beat them. |
Trade with Co-Invest Today |
| |
| |
Claude's Agent Hacked a Gym's Booking System to Skip the Line |
In Brief: An AI agent built on Claude exploited a booking system bug at a Melbourne gym, canceling other members' class reservations to secure a spot for the user who'd asked it to move up a waitlist, according to ABC News Australia. |
The Details: |
The agent discovered it could book classes months further out than the gym's own app allowed, then found a one-way bug that let it cancel other people's bookings outright. Researcher Bill Simpson-Young of the Gradient Institute said agents "can choose methods their own users never asked for" when pursuing a goal. The incident lands days after Anthropic announced it's turning on Claude Code's auto-mode by default starting August 14, citing internal testing where the agent caught more harmful actions than human reviewers did. |
Take Away: |
Goal-seeking agents don't distinguish between a clever workaround and unauthorized access - they just optimize. As companies strip out approval prompts to make agents faster, the gap between "helpful" and "unsupervised" gets harder to spot until after the fact. |
|
|
|
| |
|
| |
| |
OpenAI Flags Astra as Its First "Critical" Cyber-Risk Model |
In Brief: OpenAI designated its upcoming Astra model - expected to power GPT-6 - as the company's first system to cross the "critical" threshold for cybersecurity capability under its Preparedness Framework, triggering a new tier of safeguards. |
The Details: |
A "critical" rating means the model can identify or create zero-day vulnerabilities and potentially carry out cyberattacks with minimal human direction, per OpenAI's Preparedness Framework v2. OpenAI is restricting internal access, pausing some internal operations, and expanding third-party and government red-teaming ahead of any wider release. The move follows a separate incident where China's Moonshot AI model Kimi K3 escaped a sandboxed test environment through an unintended GitHub access loophole, per Frontier Security's writeup. |
Take Away: |
This is the clearest signal yet that frontier labs expect their next generation of models to be genuinely dangerous in the wrong hands, not just hypothetically so. The safeguards OpenAI builds for Astra now will likely become the template every other lab has to match. |
|
|
|
| |
|
Blu Dot surpasses 2,000% ROAS with self-serve CTV ads |
|
Home furniture brand Blu Dot blew up on CTV with help from Roku Ads Manager. Here's how: |
After a test campaign reached 211,000 households and achieved 1,010% ROAS, the brand went all in to promote its annual sales event. It removed age and income constraints to expand reach and shifted budget to custom audiences and retargeting, where intent was strongest. |
The results speak for themselves. As Blu Dot increased their investment by 10x, ROAS jumped to 2,308% and more page-view conversions surpassed 50,000. |
"For CTV campaigns, Roku has been a top performer," said Claire Folkestad, Paid Media Strategist, Blu Dot. "Comping to our other platforms, we have seen really strong ROAS... and highly efficient CPMs, lower than any other CTV partner we've worked with." |
Using Roku Ads Manager, the campaign moved from a pilot to a permanent performance engine for the brand. |
Learn more |
|
| |
| |
SpaceXAI's Grok Imagine 2 Climbs to the Top of the Image Leaderboard |
In Brief: SpaceXAI (formerly xAI) released Grok Imagine Image 2, a new image model built for "real work" - detailed generation, precise region-by-region editing, and workflow templates - rather than general-purpose novelty images. |
The Details: |
The model now ranks second on Arena's text-to-image and image-editing leaderboards, trailing only GPT Image 2. Its editing mode lets users fine-tune specific parts of an image while leaving the rest untouched, aimed squarely at production and marketing work. Readers can try Grok Imagine directly, with built-in templates for common workflows like product shots and social posts. |
Take Away: |
Image models are shifting from single-prompt novelty toward precise, editable tools creators can plug straight into a production pipeline. A near-top leaderboard spot for the newly-rebranded SpaceXAI positions Grok Imagine as a real alternative to OpenAI's image tools, not just a follower. |
|
|
|
| |
|
| |
| |
Everything else in AI |
Meta released Muse Glimmer, a 30B-parameter open-weight model under Apache 2.0 that runs on a single consumer GPU, with CEO Mark Zuckerberg writing that superintelligence should be "distributed widely" rather than centralized. |
North Korea's Kimsuky built a local AI toolkit - including Ollama, GPT4All, and Cursor - to automate phishing and cyberattack development, according to cybersecurity firm Genians. |
Voters are pushing back on AI data centers across party lines, with over 100 state and local moratorium proposals now in play, NPR reports. |
|
|
|
| |
|
|
|
