Court filings show ICE's investigations arm spying on left-wing activists on a scale not seen in decades.
What the article says
- Lawyers for 15 Minnesota defendants got court discovery showing a sweeping spying operation run by Homeland Security Investigations, part of ICE, against left-wing movements.
- The records include intercepted Signal chats, undercover reports and database queries, under programs called Project Whipple Shield and Operation Puppet Master.
- About 120 agents worked the case, including undercover officers who reportedly infiltrated churches despite a court order. Drones and anti-terrorism tools were used.
- Defence lawyers compare it to COINTELPRO, the FBI's old campaign against civil rights and anti-war groups. Former officials say it began well before the Minnesota crackdown.
What HN is saying
- Many were struck by how one activist was identified: agents infiltrated a Signal chat, then used cloud and payment records plus bought commercial data, with no warrant. thread ↗
- Fly.io and Stripe took heat for handing over data without a fight. Some commenters say they will stop using Fly.io. thread ↗
- The sharpest split is over ICE itself. One person asks why visa overstays should not be enforced, and replies say the problem is the force and intimidation, not enforcement. thread ↗
- Some argue ICE should not be doing intelligence work at all, since it is a civil agency. Others say mass surveillance is the real problem, because power always gets abused. thread ↗
- Several ask where the gun-rights crowd is. Replies say some are speaking up, while others say they never opposed this. thread ↗
Apple's own git and curl still accept forged certificates, and the latest macOS update made it worse.
What the article says
- Apple was warned eleven days ago, yet the git, curl and python that ship with macOS still trust the compromised eMudhra certificates.
- Anyone on your network path could tamper with clones, pulls and pushes, or steal your GitHub tokens.
- The latest macOS update did not fix it. It also wiped the manual workaround people had applied.
- Safari and apps using Apple's own networking are protected. Command line tools read a flat certificate file instead.
- Until Apple ships a fix, developers must remove the bad roots by hand and repeat that after every update.
What HN is saying
- Many say the real fault is macOS, not git. Git rightly defers to the system's trusted certificates, so the headline undersells the scope. thread ↗
- Homebrew users are mostly safe because it ships its own certificate bundle. A maintainer says git itself is still being fixed. thread ↗
- People are baffled that Apple has no quick way to push small security patches, and that its fix is a pile of commands to paste. thread ↗
- Real disagreement over the danger. Some call it overblown since browsers check certificate transparency. Others point out that a fake coffee shop wifi hotspot is enough to attack command line tools. thread ↗
- A wider gripe: certificate revocation has quietly been given up on, and Linux has its own mess of separate trust stores. thread ↗
A Russian anti-bot script hides fake bioweapon code to make AI models refuse to read it.
What the article says
- A researcher crawling the web for fingerprinting scripts found one full of alarming function names about poison doses and bomb blasts.
- It turned out to be a trick. The names, plus a hidden block of invisible text, exist to set off the safety filters of AI models and stop them from helping anyone pick the code apart.
- The hidden text is a fake jailbreak prompt covering hacking, bioweapons, self-harm and Tiananmen Square, aimed at both Western and Chinese models.
- Claude-style assistants refused to help, but an open-weight Chinese model went along with it.
- Underneath the theatre, the script is an ordinary bot detector, made by a Russian traffic security firm.
What HN is saying
- Several people noticed the same pattern: open-weight models like GLM will help with security work that the big commercial models refuse, and Hugging Face said as much after its own breach. thread ↗
- A new word was coined for this trick, "scare-crowing", and commenters traded other names for it. thread ↗
- One view is that any safeguard a model has will eventually be used against it, and that defenders will need those limits removed. thread ↗
- A skeptic says this will not hold up, because coding agents read the code structure and simply ignore dead code with scary names. thread ↗
- A customer worry: a site's brand could end up tied to bioweapons and cyberattacks in an AI's mind. Another reader replied that models do not train on obfuscated scripts, so that is very unlikely. thread ↗
- On browser fingerprinting itself, one commenter blamed browser makers. The reply was that the same features power games and drawing apps, so they cannot just be removed. thread ↗
Google announces Gemini 4 Carbon, a top-ranked model that almost nobody can use yet.
What the article says
- Google calls Carbon its strongest model, and says it tops a widely watched independent intelligence ranking.
- It is pitched at security work: finding software flaws, exploiting them and patching them, with big claimed leads on all three.
- Google says it cheats less on coding tests and fabricates less, and that it resisted every prompt injection attack it tried.
- Access starts with a small group of cyber defenders, then trusted partners, with a wider release promised only as soon as possible.
- Pricing will match the Argon model released two weeks ago, at a low introductory rate.
What HN is saying
- The loudest complaint is naming. People can't tell how Carbon relates to Argon, Pro or Flash, and Google's own apps still show older models. thread ↗
- Many think announcing benchmarks without a release date is pointless, since rivals will likely have caught up by the time Carbon ships. thread ↗
- Some doubt that any Gemini model is a real cyber risk, and suspect the limited rollout is manufactured scarcity. thread ↗
- Commenters were unconvinced by the claim of zero successful prompt injections. They say attackers aren't limited to the attacks Google tested. thread ↗
- Users who tried Argon say it is very good value, and one noted the timing, just before Google's earnings call. thread ↗
A VP of engineering charts his GitHub history to show AI made him work more, not less.
What the article says
- The author uses agents for all his engineering work and likes it, but feels busier and more drained than ever.
- The load changed shape. He used to hold two or three serious tasks a week. Now several agents run in parallel, each needing him "sometimes", and the constant switching wears him down.
- His GitHub data backs this up. Despite a leadership role with less hands-on coding, his activity is back at his busiest full-time developer levels, with August 2026 his biggest month ever.
- Cheap code means he takes on bugs, cleanups and ideas he would once have skipped, so the work expands to fill the space.
What HN is saying
- Several people say the fix is discipline: a written daily plan, a hard limit on scope, and no side quests. Replies add that the side quests land on colleagues as extra reviews and discussions. thread ↗
- The sharpest pushback is that the premise is naive. Employers were never going to pay the same for less work, so the likelier outcomes were more output or layoffs. thread ↗
- Some say AI turned satisfying mundane work into cheap, lower-quality output you now have to produce anyway, leaving only exhausting high-stakes decisions. thread ↗
- A counter view: skipping typos and log-combing leaves the fun parts, and a lead agent tracking tasks helps. Others reply that detail-lovers lose the part they enjoyed. thread ↗
- One commenter links it to Cory Doctorow's "reverse centaur" idea, where the human becomes the machine's assistant. Another says you stay in control only by running one or two tasks at a time. thread ↗