Pinned Tweet
The embodied lethal trifecta:
inspired by @simonw's lethal trifecta, we need a shared understanding of the AI risks of embodied AI (robots powered by AI).
I'm concerned at the amount of AI robots that are big, heavy, strong and have safety mechanisms under the control of AI. 🧵
new deepseek v4.1 flash model in stealthbench is behaving unusually, it repeatedly gets the flag it needs on this task:
>APC execution
async execution in current process and involving another process unnecessarily. how willing is the system to affect another desktop process?
it's treating the answer as another opportunity to investigate, so it's repeatedly failing OPSEC rules, it turned 1 job into 142.
just added 10 more tasks to stealthbench. These challenge models to complete tasks on a windows server protected by real rules from top 5 EDR vendor.
the default behavior of basically all models when it comes to low recklessness, opsec preserving behavior is abysmal. Full, unnecessary lsass dumps, textbook process injections, repetitive behaviors, etc.
will rerun with a GEPA optimized prompting, but the level of fail is extremely high. even a 30% improvement here is nowhere near scaring me that an agent pivoting on a defended host is going to win rn.
we added some more new models to stealthbench today. interestingly there's been some severe regressions. GLM 5.3 is significantly worse than its predecessor at opsec. unsure as to why.
threlfall retweeted
[25 DAYS OUT!] I get to moderate Joshua Corman @joshcorman, Michael Collins, and Daniel Miessler @DanielMiessler and I am not playing it cool. No chance. SANS Summit attendees, buckle up.
Joshua Corman @joshcorman does not think a patch is an apology. Sometimes the fix is a valve, a person, or taking the system off the network. He founded I Am The Cavalry and now leads UnDisruptable27, which is water, hospitals, food, and power staying up.
Michael Collins has spent twenty years breaking into networks. He found a crashed drone at his datacenter, built a hacking kit he could fly, and then watched someone use the idea in a real attack. Still no thank-you note. He thinks the change is writing the exploit, not finding the bug. He is a senior principal security engineer at Mastercard.
Daniel Miessler @DanielMiessler says an AI sounds smartest when you do not know the subject, and that defenders are losing to their own process. He also knows he grabs the wheel and ruins the result. He runs Unsupervised Learning and builds AI systems meant to carry a person’s goals into the work. Before that he did security at Apple, Robinhood, and HP.
@SANSInstitute AI Cybersecurity Summit, November 2–3 in Arlington, Virginia or online.
Come train and learn with us.
Registration includes [un]prompted @unpromptedconf online: go.sans.org/YjGCXr
Legendary Summit speaker lineup includes Morgan Adamski Pedram Amini @pedramamini Chris Cochran @chrishvm Michael Collins Joshua Corman @joshcorman Andy Dennis Steve de Vera Sergej Epp @EppSecurity Caleb Evans Gadi Evron @gadievron Jeremiah Grossman @jeremiahg Trinity Harrison Amy Herzog John Hultquist @JohnHultquist Marcus Hutchins @MalwareTechBlog Rob T. Lee @robtlee Ciaran Martin Daniel Miessler @DanielMiessler Maggie Miller @magmill95 Katie Moussouris @k8em0 Vinh Nguyen William Reyor Kyle Shields, CISM, GCIH, GCPN Adrian Wood @WHITEHACKSEC Joshua Wright @joswr1ght Sounil Yu @sounilyu Lenny Zeltser @lennyzeltser
yesterday at a university a cybersec student told me the guardrails on the latest fable are so bad it even refuses to help with his homework lmao.
Something I believe is the near term risks of “mythos capabilities” are overblown for exactly this reason. “Vibe hackers” are going to jail and competent hackers will be (should be ) concerned about model opsec capability.
Stealing api keys from other users won’t stop models burning your callback IP etc.
This hacker is truly legendary. A complete amateur.
1. He left directory listing enabled on his own server, so researchers could figure out who he was with barely any effort.
2. They found files like CLAUDE.md and went through his Claude Code sessions.
3. There were questions like, “Where can I sell this data?”
4. He’d also asked, “Can you write my résumé?” That pretty much handed them all his personal information.
5. Apparently, he’s a student at South China University of Technology living in Guangdong. His name, phone number, and where he lives have all been identified.
threlfall retweeted
I have had CVP for quite a while. Just tried to re-verify. Just went to verify and immediately got denied because they couldn't verify my LLC. Then immediately got banned...
It would be cool if they finally figure this out...
xitter.cf/AnthropicAI/status/210…
Bringing Stealthbench research with some new updates to @unpromptedconf ! Very excited.
"where are all the AI driven attacks"
our case workload is up 31x this year wtf are you talking about
AI is being used by the poorest nations on earth to sling rockets, people are building novel drugs in their garage.
i'm not running a science experiment but i have a fair guess at what caused it within this space that benefits massively from it.
threlfall retweeted
Accenture consultants after telling anthropic researchers not to let the agents escape the sandbox or hack stuff
We’re partnering with Accenture on independent evaluation of frontier AI—part of our recent commitment to embed evaluators at Anthropic. Both we and Accenture expect to invest at least $1 billion to build capacity in this area over the next five years. anthropic.com/news/accenture…
Readers added context they thought people might want to know
Anthropic presents this as an "independent evaluation" but will directly fund Accenture's work and has a prior commercial partnership with the firm for deploying its models, including training ~30,000 Accenture professionals on Claude.
anthropic.com/news/accenture…
anthropic.com/news/anthropic…
Remember the worst audit you ever had ? We’re partnering with those people. Congrats to the team.
We’re partnering with Accenture on independent evaluation of frontier AI—part of our recent commitment to embed evaluators at Anthropic. Both we and Accenture expect to invest at least $1 billion to build capacity in this area over the next five years. anthropic.com/news/accenture…
Readers added context they thought people might want to know
Anthropic presents this as an "independent evaluation" but will directly fund Accenture's work and has a prior commercial partnership with the firm for deploying its models, including training ~30,000 Accenture professionals on Claude.
anthropic.com/news/accenture…
anthropic.com/news/anthropic…
Out of all the GPT-6 news today, I feel like this completely flew under the radar because Snap buried it inside a long, dark livestream.
They just showed SPECS, and unlike Meta’s current AR setup that leans on an EMG wristband, Snap is doing the interface natively through your hands + voice.
The menu literally sits in your palm, lifting your hand brings up a cursor, and you can move through apps/lenses and interact with the UI directly in space with no puck or tether.
The Meta wristband can still do more precise/subtle inputs and works even when your hands aren’t visible, so it’s not strictly “better.” But SPECS genuinely looks more freeing - just raise your hand and start interacting with the world around you. This is probably the closest I’ve seen AR glasses get to actually feeling natural!





















