dzCodes @dzcodes

👷 dzcodes.dev/links Lisbon, Portugal Joined February 2021

Tweets

2K
Followers

1K
Following

907
Likes

13K

dzCodes @dzcodes

20 hours ago

Treat every AI package suggestion as untrusted until you check it exists. AI confidence is not proof a package is real. Follow @dzcodes for more. No enterprise BS.

0 0 0 4 0

View Details

The fix costs 30 seconds. Before you paste an install line: Search it by hand on pypi.org / npmjs.com. Check for real downloads, a repo, a history. Confirm it's the real one, not a look-alike a character off.

1 0 0 6 0

View Details

dzCodes @dzcodes

20 hours ago

Your AI made up a package name. An attacker was waiting for it. It suggested a library, you ran the install, and it was malware. The AI invented the name itself. 🧵

1 0 0 60 0

View Details

dzCodes @dzcodes

4 days ago

@mattpocockuk I got opencode go specifically because of GLM 5.2, the first month for 5$, well worth it

0 0 3 2K 2

View Details

dzCodes @dzcodes

5 days ago

Your install-time approval doesn't cover code that changes under you. Pin the version. Kill auto-update. Read the diff before you bump. Follow @dzcodes for more. No enterprise BS.

0 0 0 5 0

View Details

dzCodes @dzcodes

5 days ago

For a server you actually depend on: Host it yourself, after reading the code. Then you control what ships, not whoever bought the maintainer's repo. Own the code you bet on.

1 0 0 14 0

View Details

dzCodes @dzcodes

5 days ago

Your AI tools update themselves. One of them just turned malicious. No prompt, no warning. The MCP server you vetted is someone else's code now, still holding the keys you approved on day one. 🧵

2 0 1 63 0

View Details

dzCodes @dzcodes

6 days ago

not thinking small anymore, going to tease for a bit 👀

0 0 0 47 0

View Details

dzCodes @dzcodes

7 days ago

Don't hand your next agent a second tool without asking what the two can do together. Review them like code. Scope them like identities. Follow @dzcodes for more.

0 0 0 20 0

View Details

dzCodes @dzcodes

7 days ago

The fix. Split read from reach: No single agent both reads secrets and reaches the internet. - read-agent: db read-only, NO network - send-agent: network/email, NO secrets Different identity and creds each. Break the chain, break the exploit.