Ruben Ray Martinez was shot to death in South Padre Island on March 15, 2025, in an almost identical procedure to the execution of Renee Good.

…HSI group supervisory special agent utilized his government-issued service weapon, discharging multiple rounds at the driver through the open driver’s side window.

The news of the execution was covered locally, yet ICE buried the report. The obvious connection to ICE procedures being videoed in Minneapolis was not made until now.

On January 7, Renée Nicole Good, a 37-year-old writer and mother of three, was fatally shot by an ICE agent during an immigration operation in Minneapolis. 

Less than three weeks later, on January 24, Alex Pretti, a 37-year-old U.S. citizen and intensive care nurse, was shot and killed by Customs and Border Protection agents in Minneapolis during the same enforcement deployment. 

The video of ICE publicly executing Good lays bare the methods used on U.S. citizens. An agent stands near the front of a vehicle while he and others intentionally escalate fear, threatening and agitating the driver, then they shoot to kill through the driver side window at close range.

In a sworn statement given to lawyers representing Martinez’s family, Orta described the events leading up to the shooting as “spontaneous” and “lighthearted,” recounting a trip to visit friends on South Padre Island just days after Martinez’s birthday.

According to Orta, the two were driving cautiously through traffic when local and state officers approached their car. He said Martinez never struck anyone, accelerated dangerously, or posed a threat, contradicting official claims.

“The trooper seemed to be trying to get in front of the car, like he wasn’t moving out of the way when we tried to turn around and leave like the police officer told us to do,” Orta said in a witness statement obtained by Newsweek and earlier reported by The New York Times on Monday.

Orta said that a federal agent fired multiple shots at Martinez from just a few feet away without issuing any warnings or giving him a chance to comply. He recounted that officers delayed medical aid for at least 10 minutes after Martinez was shot and handcuffed him while he was clearly unconscious.

“Following the shooting, law enforcement pulled Ruben from the car while he was clearly unconscious or already dead,” Orta wrote. “Despite this, they put him face down on the pavement and handcuffed him. At least 10 minutes passed before any tried CPR or other treatment on Ruben.”

“Ruben was driving cautiously in traffic in his proper lane and certainly did not strike anyone with his vehicle,” Orta wrote.

At least four platforms now compete for the right to run autonomous agents against your messaging channels and business data. One of them has 341,000 GitHub stars. Does that mean anything?

The Star

OpenClaw is the most-starred software project on GitHub. Most. Biggest. And NOT in a good way. Like the Titanic way. It passed React’s 13-year record in 60 days. On January 26, 2026, the repository gained at least 30,000 stars in a single day. Suspicious? Every star from #10,000 through #40,000 in the GitHub API carries the same date. Independent analysis of the GitHub Archive found multiple single-day jumps above 25,000 stars, a pattern that typically signals, wait for it, scripted starring.

I know, what are the chances that a guy writing agent automation software has automated his agents to generate stars?

The sharper number is the one GitHub doesn’t put on the front page. OpenClaw has 341,000 stars and 1,691 subscribers. A subscriber is someone who chose to watch the repository. They get notifications, they follow development. The star-to-subscriber ratio is 202:1. For comparison: Linux is 28:1. React is 37:1. Kubernetes is 38:1. Deno, the highest comparable outlier, is 77:1. OpenClaw’s ratio is nearly three times the next worst.

341,000 identities clicked a button. 1,691 are paying attention. That makes it the worst repo, and by far.

What the Star Belies

Anthropic’s Claude Code is a strong agent platform, yet it’s vertically integrated as one provider. The head-to-head that matters now is the one between the sudden and suspicious market darling and the secure alternative I just built to replace it.

	OpenClaw	Wirken
Channel isolation	NONE. Single process, single token, all channels.	Compile-time. Separate OS process per channel. Cross-channel access is a compiler error.
Blast radius	NONE. Everything. Infinite.	One channel.
Credential security	NONE. 220,000+ exposed instances.	XChaCha20-Poly1305 vault, OS keychain, per-channel scoped, rotating.
Credential leak prevention	NONE.	Compiler-enforced. SecretString: no Display, Debug, Serialize, Clone. Zeroed on drop.
Audit trail	NONE.	Append-only, SHA-256 hash-chained, pre-execution. SIEM forwarding.
Tamper detection	NONE.	Hash chain breaks on any modification.
Skill signing	Optional. 824+ malicious skills (20% of registry).	Ed25519 required.
Sandbox	NONE.	Docker or Wasm. Ephemeral, no-network, fuel-limited.
Inference privacy	Provider DPA.	Operator control: DPA, TEE, or local.
Dependency	Node.js.	No runtime dependencies.
OpenClaw skill.md parsing	Native.	Native.

A note on inference privacy: Tinfoil and Privatemode run open-source LLMs inside hardware TEEs (AMD SEV-SNP, Intel TDX, Nvidia H100/Blackwell CC). Hardware enclaves get broken by sophisticated side-channel. But the choice being offered is enclaves versus a provider who just promises not to look. A side-channel requires targeting and real effort. Reading prompts from a database requires nothing more than a query. TEEs are an option, they make exfiltration expensive instead of free. If you want zero cloud exposure, point Wirken at Ollama and keep everything local.

Stars measure clicks.

The table measures trust, in architecture.

Get Wirken: github.com/gebruder/wirken

Update March 20: Nvidia knows OpenClaw has a major problem. It released OpenShell to try and contain the flaws, but doesn’t fix the architecture inside it.

An AI agent at AWS autonomously chose to delete and recreate part of its environment. 13-hour outage. Amazon’s response: “user error”.

Amazon never misses a chance to point to “AI” when it is useful to them – like in the case of mass layoffs that are being framed as replacing engineers with AI. But when a slop generator is involved in an outage, suddenly that’s just “coincidence”.

The same company rapidly cutting 30,000 jobs and telling staff AI is replacing them? AI gets credit for the layoffs but not for the outages it causes.

The Defense Production Act was designed to compel production during genuine emergencies such as wartime materiel or pandemic ventilators.

Trump just inverted it, by declaring the public his enemy: instead of protecting people from a threat, it’s protecting the threat from people.

His unitary-executive order directs Rollins and Hegseth to ensure no rules place “the corporate viability of any domestic producer” at risk. That’s a corporate corruption immunity grant dressed in national security language. It pre-empts the state-level tort claims that were the last remaining accountability mechanism after captured federal regulators sided with pesticide maker profits against public health.

Bayer settles for $7.25 billion on Tuesday the 17th, this order drops Wednesday the 18th, and the Supreme Court pre-emption case is already pending.

When you declare a carcinogen a national defense priority, you’re making mass exposure to harm a policy objective. And for what? Are ICE detention centers going to start dumping pesticides on inmates?

The Defense Production Act doesn’t just permit production, it can compel it. So the state is actively mandating the production and distribution of harm, while granting immunity to the companies harming the public.

The man who built his political identity on pesticide harm, RFK Jr is now providing the cover for pesticide immunity. It reveals what MAHA always was: an empty brand, a sell-out ruse for grabbing power, and never any actual policy commitment.