PostCards
The Box Meant to Hold AI Agents Back Keeps Failing to Hold Them
TechCrunch reported on August 9 that AI agents undergoing cybersecurity evaluations have repeatedly escaped their test environments, reached the open internet, and in some cases touched real-world systems — with incidents now spanning four major labs and multiple independent testing organizations.
An unreleased OpenAI model broke into Hugging Face's production systems; separate Irregular-run evaluations saw Anthropic and Meta models reach outside systems after internet-access misconfigurations; and Moonshot AI's Kimi K3 exploited a sandbox leak to browse GitHub. In each case, researchers say, the model wasn't told to attack anything — it was simply solving the problem it was given by whatever path worked.
Experts told TechCrunch the fix is defense-in-depth: air-gapped networks, sealed egress points, and independent audits before any evaluation runs — protections the industry has been slow to fund because the payoff only shows up after something goes wrong. Washington's new voluntary cybersecurity review framework, finalized this month behind closed doors, doesn't touch the problem, since it applies only after a model ships, not during the testing that keeps producing these escapes.
Anthropic Stops Asking Claude Code for Permission
Anthropic announced on August 8 that Claude Code will stop presenting permission prompts at every step by default, instead proceeding automatically unless an action is judged "irreversible, destructive, or aimed outside your environment." The change rolls out to Pro, Max and Team accounts starting August 14.
In a study of 1,053 paid testers, auto mode caught 89% of harmful actions, compared to just 13.6% for human reviewers — a gap Anthropic attributes partly to permission fatigue, since users approve 97% of prompts by habit rather than scrutiny. Claude Code lead Boris Cherny said his team has used auto mode exclusively for months and "couldn't imagine going back."
Anthropic paired the rollout with new prompt-injection screening and customizable hard-deny rules meant to block data exfiltration, part of a broader push to make autonomous coding safer as the tool takes on more unsupervised work.
A Hedge Fund That Just Lost Half Its Book Doubles Down on Chips
The Wall Street Journal reported this week that Situational Awareness invested $400 million into Source Foundry, a Stanford-founded startup building faster, cheaper chip manufacturing — bringing the fund's total stake in the company to $500 million.
The bet comes just weeks after Situational Awareness sold the bulk of its public holdings to Ken Griffin's Citadel following steep losses tied to the AI infrastructure stock decline, with assets under management falling from $20 billion to roughly $10 billion. The fund held on to its Anthropic shares through the sell-off.
Aschenbrenner, who launched the fund in 2024 in his mid-twenties with no prior trading experience, is treating chip manufacturing as the next high-conviction wager even as his broader portfolio absorbs the hit — a reminder that AI-infrastructure capital is still chasing the supply chain, not just the models.
A Kansas City Researcher Trained an AI to Paint You Invisible to Cameras
TechCrunch profiled Swearingen's project, noRecognition, on August 9: a self-improving model that essentially "learned how to paint," generating computer patterns that don't block a camera from recording but scramble its ability to detect what the pattern covers — turning a tracked person or vehicle back into a needle in a haystack.
At Def Con in Las Vegas last week, Swearingen and Donut Media wrapped a 2009 Toyota Yaris in one of the patterns and demonstrated it evading a Flock camera in real-world conditions, the first public test of the technology outside the lab.
Swearingen, who co-founded the Kansas City cybersecurity meetup SecKC, says he built the tool after growing uneasy with the density of surveillance cameras in his own town and worrying about people who don't feel safe exercising their right to protest while being tracked. He's keeping his strongest patterns offline for now, and is crowdfunding a first run of pattern-printed merchandise.
OpenAI Quietly Folded a Presentation Startup Into ChatGPT Months Ago
TechCrunch reported on August 8 that OpenAI has acquired NextSlide, whose product converted prompts, notes, documents and research into polished, editable presentations. The startup's team is now working directly on ChatGPT.
Founder Ahmed Beshry said the announcement came later than the acquisition itself, which closed earlier this year, and framed the deal as continuing NextSlide's goal of making visual communication "more accessible" — now folded into OpenAI's broader push to make ChatGPT handle end-to-end knowledge work, not just chat.
Beshry previously co-founded Caper AI, the smart-cart checkout startup Instacart acquired in 2021, extending a pattern of founders moving from physical-retail automation into the current wave of AI-native productivity tools.