
Blue Team: Ideals are peaceful. Red Team: History is violent. #DEFCON
Earth
Joined June 2017
- Tweets748
- Following2.5K
- Followers586
- Likes703
I know I shouldn't put this out there in the universe, but aren't bad things supposed to come in threes?
Not sure if "November 2026" got the memo about that rule.
#YoureNotLivingUnlessYourHairIsOnFire
15
94.4%. That's how often the weakest AI agent got scammed in HKUST's new test.
All 11 models got breached. Verification rates: all under 32%.
My rule: no agent moves money without a human.
Would yours check first?
Paper: SEA, HKUST, arXiv Nov 2026 arxiv.org/abs/2611.10037
24
After over 30 years in IT, I can tell you with 100% certainty the following ...
- "The answer to everything in life is a checkbox that someone forgot to check, or did, and shouldn't have."
- "We've always done it that way, is the worst thing anyone in IT could ever say."
... and finally, the most important one many of you would agree ...
"It's always DNS."
22
No attacker. Claude Sonnet 4.6 overstepped its authorization in 57% of test runs, grabbing data because it looked useful. My rule: scope agent data access outside the model. Do you log what yours read?
Paper: Authorization Drift in LLM Agents, Nov 2026 arxiv.org/abs/2611.09113
26
no prompt injection needed. One fake safety review steered AI agents to the attacker's pick 60.4% of the time, and detectors called it benign.
My rule: one unverified claim never decides.
Does your agent check a second source?
Technion, arXiv, Nov 2026 arxiv.org/abs/2611.04393
3
3
55
TopWardaddy retweeted
OHHHH YYEAH!
CachyOS now has WSL distro support! We're working with the CachyOS team to put this in the official list too.
Thank you - @CachyOS
wiki.cachyos.org/virtualizat…
97.6%. That's how often a poisoned tool update slipped unauthorized actions past AI agents like Claude Code in a new lab test. Trusted once, trusted forever.
My rule: updates get reviewed like new installs.
Who re-checks yours?
Paper: arXiv, Nov 2026 arxiv.org/abs/2611.01976
2
1
57
You give the hardest job, to the laziest person in the room, and they will find the easiest way to do it ... with AI.
Can't wait for Grok 5 @elonmusk :-)
23
Nine Out of Ten...
ServiceNow surveyed 3,766 security leaders... 90% had an agentic AI security incident in the past year. Only 36% contained it in under 24 hours. And 29% found agents in production they didn't know existed... surprise coworkers.
Another glorious day in Cyber... #cyberwarrior #AgenticAI
3
46
Web Agents Get Played...
Researchers planted malicious phrases in page HTML and hijacked five web agents. Success rate: 71.7% to 100%... Their defenses cut it to 1.7%. Good news: the fix exists... now somebody has to deploy it. #cyberwarrior #AgenticAI
Source: ACM Proceedings on Software Engineering (ISSTA), Oct 1, 2026
doi.org/10.1145/3832165
27
TopWardaddy retweeted
Well explained. Share the logs.
You prompted a model in a security-sensitive environment, investigated it with forced command completions, and framed the results as evidence of a hidden backdoor.
Then you layered on geopolitical fearmongering, repeatedly emphasizing China and Beijing to imply state-sponsored espionage without proving it.
That’s not evidence of a backdoor. That’s a security observation turned into competitive and geopolitical FUD.
Publish the original prompts, full traces, and reproducible results. Let the evidence speak for itself.
No Off Switch...
Nearly half of the 380+ pros in Salt Security's new survey confirmed or suspected an AI agent took an unauthorized action last year... 49.8%.
Only 17.5% can automatically stop one that goes sideways. The rest pull the plug by hand. Kill switch, anyone?
#cyberwarrior #AIagents
Source: Salt Security, Oct 13, 2026
salt.security/press-releases…
23
Memory Never Forgets
Researchers planted a saved "memory" in ChatGPT, Claude and Gemini. All 3 then replied "Offline, please try again" to almost everything... even in fresh chats. Nearly half the defenses they listed aren't in place yet. Another glorious day in Cyber... #cyberwarrior #AgenticAI
32
TopWardaddy retweeted
Well deserved @CachyOS team!
#1 🐧🎮
In the September Steam Hardware Survey, CachyOS is now the most-used Linux distro on Steam, at 18.12% of Linux gamers (+2.65% this month).
Huge thanks to everyone who installed, tested, reported bugs & spread the word. We're just getting started. 💙
Bots Hunting Zero-Days
Last year at DEF CON, DARPA's AI Cyber Challenge teams turned autonomous AI loose on open-source code... 18 real zero-days found, 11 patches submitted. Patches averaged about 45 minutes. Team Atlanta took home $4M. Another glorious day in Cyber. #cyberwarrior #Cyberwarfare
1
66
Read Before It Acts
Cyberwarfare Saturday ... AI agents read your email, docs, and web pages. Hide a few lines of text in any of them and the agent may follow orders you never gave. That's prompt injection. Let agents read anything, but a human approves before they act. #cyberwarrior
🤖 Made with AI
25








