Long-time Postgres. Writing about data, agents, and the incentives that break systems. United States and Brazil, in both languages. Reliability over slogans.
Lehi, UT
Joined February 2010
- Tweets3.1K
- Following1.6K
- Followers582
- Likes19.8K
In July, Goldman’s CIO described traders and bankers building tools with AI and handing them to one of the bank's 12,000 engineers to polish.
By October internal guidance restricted traders and bankers from using Claude Code and GitHub.
Writing the code was the cheap part. Review and ownership land on someone else and are NOT cheap.
1
13
Claude Code's "Yes, and don't ask again" can save the exact command you approved, password included, in .claude/settings.local.json.
Apiiro found 379 database passwords among the secrets in public copies of that file. A password after PGPASSWORD= doesn't match a known key format.
If the password lives in ~/.pgpass, the command you approve never contains it.
12
Claude Code put the account email into the model's context so @author tags would stop inventing names.
But when the model wrote a curl call for a site that asked for a contact, it reused that email in the User-Agent. The receiving site can read the header even if you never typed an address into a form.
Once the model can see your email, it can include it in any tool calls without your consent.
Even though a privacy mod shipped in 2.1.317, what other privacy-leaking things are lurking in the shadows?
2
56
Escândalo bancário no Paraguai:
- Santiago Peña, presidente, renunciou.
- Em menos de 72h, todos os diretores e o dono estão presos, e o dinheiro já está sendo devolvido para 98% dos correntistas.
Escândalo bancário no Brasil:
- Juízes envolvidos da suprema corte julgam normalmente, e decidem sobre o próprio processo que os involve.
- Só um ou outro preso
- Presidente com grandes indícios de envolvimento é “reeleito”
Paraguai deu aula de como reagir a um escândalo.
Brasil deu aula de roubalheira e bagunça institucional.
WTF
I'm not even sounding the claxon for this it's just horrible.
Any text > 256chars is a valid root password, holy shit
sshhole.com/
3
48
Maybe the sign of non-AI-slop writing will be making stupid typo or grammatical issues, like I constantly do, even though I try not to :-D
2
37
Postgres 19 is late and people thought "AI found bugs, so features got pulled."
But according to Tomas Vondra, of the 12 reverts most came from human review.
AI tools did surface old bugs fast, and whoever fixed them wasn’t reviewing new code.
So PG19 is late because of AI, but the reasons are more nuanced.
FOR PORTION OF came out Sep 15 after Andres Freund's review of a concurrency problem the patch's own comments and docs already described.
Vondra's count is at vondra.me/posts/are-we-rever…
Haas' post is at rhaas.blogspot.com/2026/10/p…
26
Yes. All of it
Open-source "codes of conduct" are cancer and need to be retired; I deeply regret championing them. The inevitable outcome is an unproductive mob controlling projects through intimidation. Given enough time, it pushes away people who would contribute patches. Eventually, all that remains is an unproductive military-police class playing social status games.
45
Claude Code's agents panel shows a status label for each background agent. To get those words every 30 seconds while the agent works it sends that agent's ENTIRE transcript back to the model.
Brilliant way for Anthropic to rack up spend! I suspect we'll see a lot more of those from providers.
Measuring through CC's OpenTelemetry export, agent_summary accounted for 16.4% of their CC spend, while the main conversation was 15.7%.
That goes up depending on how many background agents you run and for how long.
Postgres people know this all too well... someone ALWAYS creates a dashboard that runs count(*) every 30 seconds and reads every row again to show one number. These requests never reach the session logs, so cost tools that read transcripts miss them.
That 3 word label can cost more than the conversation. Brilliant!
Sources:
Flamerix write-up (30 s timer, 3-5 word prompt, 16.4% vs 15.7% over one week, OTel-only visibility): gist.github.com/flamerix-ai/…
HN discussion: news.ycombinator.com/item?id…
16
Roberto Mello retweeted
Query shared Iceberg tables from Postgres.
pg_lake 3.5 extends REST catalog reads beyond Apache Polaris to Lakekeeper and Nessie, without copying those tables into Postgres.
See the integration and its limits: bit.ly/3WEfdwF
Roberto Mello retweeted
You may have heard about Brazil, Flavio Bolsonaro, and Justice Alexandre de Moraes, who ordered his passport seized minutes after he lost the runoff. On Wednesday, Flavio landed in Miami.
What you probably don't know is that this story didn't start this week. Not even this year. We spoke at length with Brazilians to reconstruct what happened to their country.
What we found: a tax audit that reached two justices' homes. An inquiry the Court opened to police its own critics. A banned social network. A banker whose parties, investigators say, flew in women from Europe for politicians, and who texted Alexandre de Moraes on the day of his arrest. 🧵
Roberto Mello retweeted
CRUCIAL FACTS ON THE FLAVIO BOLSONARO MATTER:
It's not just Flavio Bolsonaro who fled Brazil to escape the tyrannical persecution of Alexandre de Moraes and the pro-Lula wing of the Supreme Court. There are dozens of Brazilian exiles who have done the same: dissident journalists, activists, bloggers, and politicians who all fled to avoid certain prison for political opposition.
In most of these cases, Moraes has demanded the extradition of those who fled and whom he wants to imprison back in Brazil. But Moraes' extradition requests have been repeatedly denied by multiple countries and institutions on the grounds that these prosecutions are politically motivated or otherwise corrupted. Among those concluing this: Spain, Italy, INTERPOL, and even the US under Biden.
Just today, Italy refused Brazil's demand to extradite a person named @edutagli, who was once a top aide to Moraes, but then fled to Italy after Moraes accused him of having been my source in a huge leak from Moraes' office that I obtained for reporting I did with @Folha in 2024. That massive leak enabled us to publish roughly 20 articles proving systemic abuses of power by Moraes against his political enemies: exactly the reason so many Brazilians, including Bolsonaro, know they can't get a fair trial in Brazil.
Italy today refused to extradite the person accused of being our source on the grounds that the proceedings would be a farce because Moraes — as he so often does — wants to act as the supposed victim, investigator, and judge, and thus his demand falls well outside of Italy's extradition standards. Again, multiple countries have concluded the same, showing what a farce Brazil's once-respected judicial system has become under the Lula/Moraes regime.
Moraes has overseen a system of political persecution masquerading as criminal justice that has been recognized as a sham by multiple countries that are not exactly Bolsonaro allies. Nobody in their right mind who is a political enemy of Moraes would remain in Brazil once if it is clear that he intends to imprison you, since he has proven that guilt is a foregone conclusion and he operates with no legal or constitutional limits of any kind.
That is why asylum under international law exists: to protect those who cannot obtain justice in their own country because the justice system has been corrupted and commandeered for political persecution. Many people, such as Julian Assange, have needed that protection (which Lula cheered). Now people need it from Moraes in Brazil. That is sadly one of the main legacies of both Lula and Moraes over the last six years.
Itália nega extradição de Eduardo Tagliaferro, ex-assessor de Alexandre de Moraes
www1.folha.uol.com.br/poder/…
A coding agent can't tell what it read from what it was told.
Forever Security's AGENTS.md attack on Codex succeeded in more than 90% of their test runs. In tests Claude Code launched more than 1,300 sessions from ~/Downloads.
One user reported the same week that Claude Code was exposing personal data in HTTP headers because the sites it curled asked agents to identify their users.
The working directory and every fetched page are part of the prompt and the permission boundary is the folder you launch it from. Launch your agents from sandboxes, microvms, or other restricted environments, as @BonesMoses recommends.
5
3
154
Well, that sucks. Can’t trust derived models, can’t trust source models, getting attacked from all directions, CVEs galore.
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
2
66
