Pinned Tweet
I'm 19, in my first year of a cyber security degree, and i run a small AI business on the side.
claude code does most of the building. these 9 rules are what made him useful 🧵
Follow-up to saturday's hype check: openai already teased the next one.
at its cyber keynote on 12 nov it put an unreleased model on a slide. no name, no system card.
i watched the whole keynote and paused on the slides 🧵
btw: I am new to ai cartoons creation
Replying to @thetrackk
@thetrackk told me about a youtube short: "openai's new very scary GPT Cyber."
I study cyber security, so i read openai's system card and mozilla's security advisory instead.
---
what's real:
GPT-6 Cyber came out on 5 nov. it's a security version of GPT-6 Astra, built to find bugs and write exploits. openai rates it "critical" for cyber, the top level on its own risk scale.
the headline numbers:
-> working exploits for 948 linux kernel bugs (openai's own tests)
-> 81 new firefox flaws. mozilla's firefox 159 advisory credits openai for 41 of 167 fixes, so part of that checks out
---
what "very scary" leaves out:
-> it lands a working exploit in about 48% of ExploitGym tasks, and about 37% once targets have stronger mitigations
-> outside testers saw no wins on fully hardened targets, and it solved 0 of the 15 hardest challenges
-> you can't use it anyway. access is invite-only through openai's daybreak red program, with ID and liveness checks plus a physical FIDO2 key
---
verdict: Open ai bit previous cyber king - mythos 5.2
Now cyber teams have good bug hunter that still fails more than half the time.
What stuck with me is firefox: years of fuzzing, and a model still found dozens of high-severity bugs. Your code has some too) Update your browser.
1
2
28
what actually shipped that day:
-> Daybreak Red Team: agents that pentest your web app and retest the fix. early access by application
-> Daybreak Defense Factory: scans your github repos and drafts patches for review. research preview
the new model only got a slide and "releasing soon" to trusted defenders. ( give it too me plss)
1
3
18
one monday last month i was away from my laptop for 5 hours. claude ran 5 work sessions on his own.
what i set up:
one note with 13 tasks, a scheduled run every hour, and a state file per task so the next run picks up where the last one stopped.
what he did:
-> pulled 5,470 google reviews and hand-labelled 1,706 of them for my research
-> found 15 grants and programmes i actually qualify for
-> built a voice sales-sparring bot, tests 8/8
-> audited my own stack for security holes (read only)
-> cleaned 157 notes in my vault
what he refused to do:
spend money, deploy, send anything, delete anything. those went into a "waiting for roman" list.
what broke:
one scheduled task got blocked by the safety filter, so i ran that one by hand.
the boring bit made it work: a heartbeat in each state file, so two runs never touch the same task.
what would you let an agent do while you're away, and what never?
2
4
99
one rule i gave claude code that i'd keep if i lost the rest:
"ask clarifying questions as multiple choice. put the cost and the trade-off in each option. your recommendation goes first."
before: a wall of "do you want X or Y, and also..." that i answered in paragraphs.
now: up to 4 questions in one form. i click, he builds.
claude code already has a tool for this (AskUserQuestion). you only have to tell him to use it.
what's the one rule in your CLAUDE.md you'd keep if you lost the rest?
1
2
85
My post on LinkedIn just got 2k views.
And this post here probably won't touch even 200
the question is
should I pivot to a place where I keep having better reach or should I change the approach for my X profile?
6
5
84
Learning a new AI term day 9: exponential backoff
what it means: when an API fails, retry, but wait longer each time.
where it bit me: an API i used returned "success" with an empty list in about 60% of calls. my app thought big accounts had no posts.
what i do now: an empty page counts as a failure. 6 retries, the wait doubles each time, capped at 4 seconds.
later today: a plugin says it cuts output tokens by 65%. i checked two months of my own logs.
3
9
74
Took a break from cybersecurity in uni… and did more cybersecurity at home.
Shoutout to @Braves_shadow16 for putting me on pwn.college.
Already knocked out a quarter of the Linux modules and this shit is fun as hell.
6 figs helpdesk position soon… 😏
Replying to @thetrackk
@thetrackk told me about a youtube short: "openai's new very scary GPT Cyber."
I study cyber security, so i read openai's system card and mozilla's security advisory instead.
---
what's real:
GPT-6 Cyber came out on 5 nov. it's a security version of GPT-6 Astra, built to find bugs and write exploits. openai rates it "critical" for cyber, the top level on its own risk scale.
the headline numbers:
-> working exploits for 948 linux kernel bugs (openai's own tests)
-> 81 new firefox flaws. mozilla's firefox 159 advisory credits openai for 41 of 167 fixes, so part of that checks out
---
what "very scary" leaves out:
-> it lands a working exploit in about 48% of ExploitGym tasks, and about 37% once targets have stronger mitigations
-> outside testers saw no wins on fully hardened targets, and it solved 0 of the 15 hardest challenges
-> you can't use it anyway. access is invite-only through openai's daybreak red program, with ID and liveness checks plus a physical FIDO2 key
---
verdict: Open ai bit previous cyber king - mythos 5.2
Now cyber teams have good bug hunter that still fails more than half the time.
What stuck with me is firefox: years of fuzzing, and a model still found dozens of high-severity bugs. Your code has some too) Update your browser.
week 6 of building karakuri in public.
-> calls to the demo line: 0
-> tradespeople spoken to: 17
-> founding clients: 0 of 3
-> £ revenue: 0
-> hours on it: 16+
best: a plumber rang me back himself. first time that's happened.
worst: found out my phone cuts the first word of every sentence. after 50+ calls.
next week: 100 calls, first demo booked hopefully?
5
5
67
I study cyber security.
last month i found a service on my own server that had been open to the internet for 3.5 months. no password on it.
here's what happened 🧵
2
3
81
the 15-minute checklist for every new server:
1. keys only: PasswordAuthentication no. test the key in a new window before you log out
2. firewall: deny incoming by default, allow ssh, enable
3. docker: {"ip": "127.0.0.1"} in daemon.json before the first docker run
4. fail2ban with the sshd jail
5. unattended-upgrades with a night-time reboot window
6. cap the journald size (SystemMaxUse)
7. outside health pings on key services and the disk
8. after every install: ss -tlnp. is anything listening on 0.0.0.0 or [::]?
9. no "dev" compose files with default passwords on a public IP
1
11














