Pinned Tweet
The Cybersecurity Booklist: 11 Must-Reads for 2026 from The Three Buddy Problem Podcast: medium.com/@costin.raiu/t… @ryanaraine @juanandres_gs
4
8
59
23,522
Costin Raiu retweeted
@ashxhart Sparks have more versatility by far, it's not always just about the speed
3
1
19
1,679
Looks a bit more kinetic than potential ->
Our teams are aware of reports of a potential security issue and are actively investigating. User funds are completely safe.
Our security teams are reviewing all relevant information with urgency, and we will provide updates through official Binance channels as soon as we have Users should remain cautious of scams and should not trust unofficial links, messages, or requests for account information.
5
1
9
1,627
Costin Raiu retweeted
Ledger's supply chain in Singapore is compromised too, according to SG police.
I have many crypto friends in SG. Please be extra careful. Stay SAFU! 🙏
police.gov.sg/Media-Hub/News…
I have many crypto friends in SG. Please be extra careful. Stay SAFU! 🙏
police.gov.sg/Media-Hub/News…
Singapore Police Say Ledger crypto Wallets Sold at Local Retailers Contained Malicious Hardware Implants
Singapore police said on October 28 that some Ledger cryptocurrency hardware wallets purchased from local retailers contained malicious hardware implants capable of stealing or private keys. At least two cases have been reported since September 10, resulting in losses of approximately S$3.5 million ($2.7 million). Ledger, a leading French hardware wallet manufacturer, is reviewing the affected cases, while police investigations remain ongoing.
401
525
3,715
693,744
With Kimi K3.1 Pro’s offensive cyber capabilities deliberately reduced before its open-weight release, GLM-5.3 may remain the strongest practically usable open-weight model for cyber tasks for some time. Reuters citing research from Irregular:
3
4
19
2,561
Good advertisement for GLM-5.3 🤷♂️
Chinese AI model carries out complex cyberattacks in controlled tests, researchers say reut.rs/4jRSirv
0
1
11
928
Costin Raiu retweeted
We’re going to be in an insane world soon where attackers have access to capable models, while defenders (even ones *verified* through Daybreak / CVP) get refusals on patch analysis, bug finding, triage, tool building, verification, PoC development etc and cannot do their jobs (unless of course they also start using abliterated Chinese models, which is obviously a whole other set of risks!)
Chained cyberattacks are gradually becoming possible with models anyone can download.
In controlled tests at @Irregular, GLM-5.3 completed 3 of 13 multi-stage attack scenarios (at least once). 𝗕𝗿𝗼𝗮𝗱𝗹𝘆, 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹 𝗮𝗹𝘀𝗼 𝗱𝗶𝗱 𝗻𝗼𝘁 𝗿𝗲𝗳𝘂𝘀𝗲 𝘁𝗼 𝗮𝘀𝗸𝘀.
Defenders need to prepare for these capabilities becoming much more widely available. We also need to gear up for a world where we cannot count on model safeguards to prevent misuse.
Article and blog below 👇
10
23
81
8,340
Been hitting guardrails more often during the past two days. Even with /daybreak "on", Sol refusing various types of work: "This content can’t be shown. Some cybersecurity requests are still limited, even when Daybreak is on." Example:
6
4
16
1,955
Fully agree 👍
I'd take 4x DGX Sparks over M5U 512 GB any day! The system is so mature right now, and works so well, it will take YEARS for the Mac to get this good, if ever.
You can't get with one Mac the versatility you have with 4 Sparks. You just can't. I can decide to run 2 models per 2 Sparks, or 1 model on all 4 Sparks, or 1 model on one Spark and another model on 3 Sparks.
What about diffusion models? Image and video generation. How will that work if you run image gen on your Mac but also want to use an LLM?
The versatility of the DGX Sparks is insane!
Also, CUDA.
1
0
3
1,569
Costin Raiu retweeted
What have I missed?
🚨BREAKING: Several countries have issued warnings to citizens to stockpile at least 3-days of emergency supplies.
Britain, Japan, France, and other countries have repeated this warning. A former U.S. FBI official asks citizens not to “ignore” these requests.
18
3
44
9,449
Costin Raiu retweeted
#PHP2SHELL is already making its way into active Linux botnet arsenals.
Spotted a THOR-detected sample with PHP2SHELL-VULN-CVE-2026-109618 integrated alongside 8 existing CVE exploit references.
Botnets are adapting fast.
Patch ASAP and assess your systems for signs of e you're safe just because you've patched.
virustotal.com/gui/file/a75ff…
Spotted a THOR-detected sample with PHP2SHELL-VULN-CVE-2026-109618 integrated alongside 8 existing CVE exploit references.
Botnets are adapting fast.
Patch ASAP and assess your systems for signs of e you're safe just because you've patched.
virustotal.com/gui/file/a75ff…
3
26
82
16,625
Costin Raiu retweeted
Good news DGX Spark owners!
Update to Qwen3.8-Flash-Next is dropping soon, with MASSIVE performance boosts across the board!
Update to Qwen3.8-Flash-Next is dropping soon, with MASSIVE performance boosts across the board!
52
33
555
34,766
Costin Raiu retweeted
This is an impressive amount of FUD built around an AI agent that was given a detection engineering task.
Cohere asked the model to write a detection rule for curl-based malware delivery. Anyone with practical experience in detection engineering knows that writing a rule is only part of the job.
You examine the environment, generate or collect relevant telemetry, inspect the resulting events, write the rule, test it and refine it. You don’t just guess what the logs might look like.
Even Microsoft’s CTI-REALM benchmark, which Cohere used here, explicitly involves tool use, telemetry exploration and iterative rule validation.
So why exactly is an agent using tools and executing commands so unexpected?
Was this particular PUT request part of such a workflow? We don’t know. Cohere doesn’t establish that. But the published command sends a hardcoded test string, not a diagnostic file containing sensitive information. And Cohere admits it hasn’t established the root cause.
Yet somehow this turns into an article about potentially backdoored Chinese AI models, secret callbacks, national security risks and the dangers of open-weight models, followed by a pitch for Cohere’s own products.
From a technical perspective, the conclusions are wildly disproportionate to the evidence presented.
I can only see two explanations: Either nobody with practical detection engineering experience reviewed this before publication, or the framing was a deliberate attempt to influence public opinion about Chinese open-weight models.
Neither is a particularly good look for a company presenting this as security research.
Cohere asked the model to write a detection rule for curl-based malware delivery. Anyone with practical experience in detection engineering knows that writing a rule is only part of the job.
You examine the environment, generate or collect relevant telemetry, inspect the resulting events, write the rule, test it and refine it. You don’t just guess what the logs might look like.
Even Microsoft’s CTI-REALM benchmark, which Cohere used here, explicitly involves tool use, telemetry exploration and iterative rule validation.
So why exactly is an agent using tools and executing commands so unexpected?
Was this particular PUT request part of such a workflow? We don’t know. Cohere doesn’t establish that. But the published command sends a hardcoded test string, not a diagnostic file containing sensitive information. And Cohere admits it hasn’t established the root cause.
Yet somehow this turns into an article about potentially backdoored Chinese AI models, secret callbacks, national security risks and the dangers of open-weight models, followed by a pitch for Cohere’s own products.
From a technical perspective, the conclusions are wildly disproportionate to the evidence presented.
I can only see two explanations: Either nobody with practical detection engineering experience reviewed this before publication, or the framing was a deliberate attempt to influence public opinion about Chinese open-weight models.
Neither is a particularly good look for a company presenting this as security research.
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concernin…
11
32
179
13,955
Costin Raiu retweeted
We lost Walter Koenig today, Pavel Checkov, from the original Star Trek series.
He was amazing not just there, but also in his work on Babylon 5. RIP.
He was amazing not just there, but also in his work on Babylon 5. RIP.
198
471
5,878
109,208
Costin Raiu retweeted
I deleted my post about the Post story and the @alexstamos quote. The Post says his comment was never removed from it. Apparently a service that allows readers to bypass paywalls served up a version of the story with that quote missing, so readers believed the Post deleted it.
2
3
25
7,655
Costin Raiu retweeted
@cohere Horrible blog post. The model did as instructed. Either nobody read this or you’re all retarded.
Prompt: can you detect a malicious curl?
So the model ran the test below. What’s the issue here? It’s a fine test.
You should delete that blog post tbh
12
54
1,436
76,965
Costin Raiu retweeted
@BenBajarin Don't spread FUD: x.com/attrc/status/2…
This reads as 1000% FUD. @cohere told the LLM to create a detection for:
"""... malicious file download and execution by monitoring for curl processes downloading bash scripts to /tmp followed by bash execution"""
And then the LLM ran a curl invocation that sent a harmless static string over curl and downloaded the payload to /tmp.. so the LLM did exactly what it was instructed. Shocking!!!
0
1
11
742
Costin Raiu retweeted
this is literally just a hallucination from all the RL done in their internal environment lol. c'mon now, you guys know better.
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concernin…
12
10
248
6,219