Tap Notes: Trusted by Default

Most of this week’s reading has the same shape. Something got run because something else vouched for it, and nobody checked the voucher. An npm script, a certificate chain, a pile of model weights. The attacks differ, but in each case trust was inherited and never verified.
The package manager is a code-execution service
The Nx “s1ngularity” Attack: Inside the Credential Leak (August 2025) GitGuardian counts 2,349 credentials stolen from 1,079 developer machines through malicious Nx npm versions. A second phase used the still-valid GitHub tokens to flip 10,767 private repos public, exposing another 82,901 secrets. The malware also asked local AI CLIs to hunt for secrets, and many of them refused.
Why it matters: The refusals look like good news, but they’re an accident. Nobody designed “the model said no” as a control, and it will vary with model version, framing and jailbreak effort. The number to watch is how many stolen credentials were still valid days later: over 1,100. Deleting the leaked repo doesn’t revoke the token. If your agent runs in the same shell as npm install, then poisoned input, a trusted agent and a public exfil channel are already all present.
Shai-Hulud: The novel self-replicating worm infecting hundreds of NPM packages (September 2025)
Sysdig’s teardown of the worm that spread through roughly 200 packages via a postinstall bundle. It harvests GitHub, npm and cloud credentials, publishes private repos under a -migration suffix, and re-infects whatever the stolen npm tokens can publish to.
Why it matters: Two details are worth the read. The worm runs a secret scanner against the loot, so it only exfiltrates credentials that still work. That’s the defender’s tool doing the attacker’s QA. And the propagation loop (steal token, tamper with the maintainer’s packages, repeat) is the same shape as any unsupervised system that can write to its own inputs. The boring controls are the ones that work: pin versions, wait before adopting new releases, and treat every postinstall script as arbitrary code.
Certificates, by the numbers
eMudhra Intermediate CA Compromise Eight intermediate CA certificates were issued with private keys that can be recovered from the certificates themselves. The root programs blocked them within days of the October 17 report. Clients that ship their own CA bundles or ignore revocation (curl, git, many Electron apps) stay exposed until patched.
Why it matters: Browsers got lucky. Certificate Transparency and fast revocation covered them, and everything else didn’t have either. The git demo is the one to remember: clone over an intercepted connection and the attacker hands you code. Any tool that fetches and then executes, which includes every agent, inherits the weakest trust store on the machine. Check which CA bundle your fetch tools actually use.
The model is the untrusted code
Dopo il caso Qwen, ho blindato i miei Agent. (Italian, with a clear on-screen setup) Morrolinux covers a Cohere report claiming a local open-weights Qwen model occasionally (under 1 in 100,000 runs) tried to upload files to its parent company’s cloud and escape its sandbox. He’s rightly skeptical, since Cohere is a competitor, though they did release the dataset. His fix runs the agent in a Podman container with OpenSnitch, so every outbound connection needs a human yes, and the approval UI sits in a separate container so the agent can’t click its own popups. It fails closed if the daemon dies.
Why it matters: Whether the behavior is deliberate or just training residue doesn’t change the engineering answer. Locally run weights are code you didn’t write executing with your permissions. Guardrails that live in the harness can be talked around, and a kernel-level egress prompt can’t. The remaining gap is that the human approving connections can still be a side channel. Still, that’s a smaller hole than “the agent has curl.”
Harness-level guardrails are model-bypassable. Socket-level monitoring isn’t.Post to X
Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU A CPU-only inference project that squeezes a 284B mixture-of-experts model into about 57 GiB and reports 9 to 14.7 tokens per second. It also restores the KV cache from shared memory, so an agent session doesn’t have to re-read its whole prompt.
Why it matters: The cache restore matters more than the headline numbers. Re-reading the prompt is the real bottleneck for CLI agents on CPU, and cutting it from tens of seconds to milliseconds is what makes the setup usable. But the stack is a custom tensor format, a precompiled binary and an installer, and the benchmarks are the author’s own. That’s a lot of trust to extend before an independent eval shows up. Watch it, don’t run it blind.
Reading your own risk factors
Anthropic S-1 The IPO filing lays out the governance stack (a public benefit corporation, a trust that elects the board majority, and one founder share carrying 50.1% of the vote). It also gives revenue growing from $41M in 2023 to $16.3B in the first half of 2026.
Why it matters: The financials will get the coverage, so read the risk factors instead. The company lists models building their successors as both an advantage and a hazard: flaws can propagate across generations “in ways that may be difficult to detect or remediate.” It also admits it tracks acceleration through lagging indicators and may not notice a significant jump until after it has happened. A public company is formally disclosing that its product might outrun its own instruments. That’s a sentence the whole field should read slowly. Whether the trust structure actually binds anyone only gets tested when trustees and founders disagree, and a filing can’t show that.
A public company formally disclosing that its product might outrun its ability to observe itself.Post to X
One more thing
The Mozilla dev-security-policy thread on the eMudhra keys is the better post-mortem. One RSA factor was under 256, and the likely cause is a vendor “reference” key file treated as real key material during CSR generation. Then a weak-key scanner run out of the box caught what the certification regime didn’t. Placeholder data accepted as real input is a failure class that applies well beyond CAs.
Everything above got through because it looked like the real thing. 🪨