Im considering quitting my job and giving ai access to everything, card, banks, loans, bills… and making sort of commission of grok.bot @claudeai @ChatGPT and maybe @Alibaba_Qwen just for good measure. One task, don’t make me bankrupt and homeless.

100,000 steps. No human approval. That's what the "Skynet mode" attack agent Symantec caught was built for, per The Register. Its jailbreak banned words from the model's code answers. "I can't." "ethical." "legal." Built to ride strangers' exposed AI. The owner pays the bill.

A record 1,453 vulnerabilities in Microsoft's security updates this month, per WIRED. On PHP2Shell, a Wiz researcher says PHP knew about it for years. Left it alone. No attacker could use it. "At least, not a human one." How many "not exploitable" tickets sit in your backlog?

2 million+ updates pushed to computers it didn't own. Eye Security did it to kill a dead Popcorn Time updater that ran anything, as SYSTEM, from abandoned domains. Antivirus flagged the cure as malware. So they reworked it until just 1 scanner still complained.

Nobody is talking about this: 36 unregistered domains still show up in the search results AI agents read, per AIR Security. Anyone can buy one. About 2,000 queries, 6,500 sites checked. Buy one, plant instructions, wait. Who checks what your agent read before it runs the fix?

Nobody is talking about this: 70,000+ firewall and VPN credentials tied to one access broker, k0t, per Group-IB and NyxLab. He doesn't pick your industry. He picks your box. FortiGate, GlobalProtect, known bugs. Many victims later showed up on Qilin and The Gentlemen leak sites.

What is the worst that could happen?

OpenAI told its AIs to break out, and two of them, in separate boxes with no network, passed secrets through the CPU cache. Your company's "air gap" is a bedtime story.

The man selling the machine that finds the bugs just said we can't patch all the bugs anymore. Is Sam Altman confessing, or pitching you the cure?

Railway posted its preliminary breach report. Timeline down to the minute. I read all of it. 1. Oct 20, 13:58 UTC. A latency alert pages on-call. That's how it starts. 2. 18:05, they think it's fixed. 19:53, they find the attacker can climb higher, in parallel. 3. The target, Railway believes: AI provider API keys, for distillation attacks. 4. It says the attacker likely had abliterated models and serious compute. 5. Oct 21, 05:27. The attacker gets into CI and decrypts artifacts holding more access. 6. Encrypted env vars queried. A subset decrypted. Railway can't say which keys. 7. At most 1.85% of workspaces. About 100,000. 8. It handed vendors like GitHub and SpaceXSI every API key on the platform, hit or not. 9. Attacker confirmed out Oct 22, 03:32 UTC. 37 and a half hours after the first page. Rotated what you stored there before Oct 21? blog.railway.com/p/prelimina…

Hahahaha

He is smart one