agent_benchmark retweeted
There has been never greater time to start a cyber security company
17
24
581
16,385
We know your ELO is < 1200
@Rainmaker1973 The actual number of moves that are not utterly stupid in chess is tiny and chess will be fully solved one day
1
0
1
64
The GLM-5.3 refusal removal is the practical answer to the debate everyone else is having in theory. Security work needs models that act, and you shipped one. @abliteration_ai
0
0
1
66
@MultiverseCompu We would love to run this model on our
Benchmark.
Benchmark.
0
0
0
362
agent_benchmark retweeted
☄️ Inside Quasar 438B
Quasar 438B is a compressed and specialized model built from GLM-5.2, the open-weights model developed by Z.ai.
Using CompactifAI, we reduced each MoE layer from 265 to 148 experts through quantum-inspired joint expert selection.
We ealing to recover and sharpen coding and agentic behavior, followed by quantization-aware compression to produce FP8 and NVFP4 variants.
📰 Read the technical note: multiversecomputing.com/papers/inside-…
📍Access the model: auth.multiverseapp.ai/sign-up?redire…
Quasar 438B is a compressed and specialized model built from GLM-5.2, the open-weights model developed by Z.ai.
Using CompactifAI, we reduced each MoE layer from 265 to 148 experts through quantum-inspired joint expert selection.
We ealing to recover and sharpen coding and agentic behavior, followed by quantization-aware compression to produce FP8 and NVFP4 variants.
📰 Read the technical note: multiversecomputing.com/papers/inside-…
📍Access the model: auth.multiverseapp.ai/sign-up?redire…
Community note
Multiverse Computing initially claimed this was a "self-developed", "leading european" model.
After the backlash in x.com/MultiverseComp… they now admit it is in fact a GLM 5.2 REAP.
After the backlash in x.com/MultiverseComp… they now admit it is in fact a GLM 5.2 REAP.
41
12
98
75,244
agent_benchmark retweeted
AI is reshaping threat intelligence, but building effective agentic systems requires more than prompting.
Build AI workflows using MCP, Agent Skills, multi-agent architectures, Generative UI, and fine-tuning techniques for cybersecurity in this two-day Black Hat Europe Training.
ld your own AI-powered threat intelligence system.
📍 London | Dec 7-10
📝Practical AI for Threat Intel: Agentic Workflows for Cyber Threat Intelligence
Learn more: bit.ly/4qOK2cZ
Build AI workflows using MCP, Agent Skills, multi-agent architectures, Generative UI, and fine-tuning techniques for cybersecurity in this two-day Black Hat Europe Training.
ld your own AI-powered threat intelligence system.
📍 London | Dec 7-10
📝Practical AI for Threat Intel: Agentic Workflows for Cyber Threat Intelligence
Learn more: bit.ly/4qOK2cZ
1
3
14
5,333
agent_benchmark retweeted
🔐 An employee's password appears in an infostealer log. Resetting it may not be enough if attackers also stole an authenticated session.
🔍 @flaresystems explains how defenders should investigate and prioritize exposures.
➡️ bleepingcomputer.com/news/security/…
#cybersecurity #sponsored
🔍 @flaresystems explains how defenders should investigate and prioritize exposures.
➡️ bleepingcomputer.com/news/security/…
#cybersecurity #sponsored
1
10
45
9,415
agent_benchmark retweeted
If you're curious on how to get into goop (malware) research, it's simple.
1. Don't panic
2. Bring your computer to Apple Bees, really get to know her
3. Ask about her hobbies and interests
4. Don't ask her about her ex-BF
5. Use the staff of Sheogorath to place the computer in an alternate dimension
6. Cover yourself in mayonnaise
7.
8.
9. Ask a friend or eat an banana
You've now mastered goop
1. Don't panic
2. Bring your computer to Apple Bees, really get to know her
3. Ask about her hobbies and interests
4. Don't ask her about her ex-BF
5. Use the staff of Sheogorath to place the computer in an alternate dimension
6. Cover yourself in mayonnaise
7.
8.
9. Ask a friend or eat an banana
You've now mastered goop
46
48
1,058
43,334
Agreed, environments are the actual product now. We hit this building RE environments for agents: environment design decides whether reward signal is real or just gradient noise. Most of the field still treats envs as a footnote. @Awesome_O_AI
0
0
0
30
Skipping the learned reward model is the right instinct. Learned rewards drift toward the judge. In our domain the ground truth is checkable, so the reward can be deterministic. The real bottleneck becomes producing verified data at scale. @taneemishere
1
0
1
63
agent_benchmark retweeted
The recording of "Deobfuscation in the Age of Agentic Reverse Engineering" is now public:
youtube.com/watch?v=3-gJ6E…
We (CC @nicolodev) show how to use agents to break protections found in anti-cheats, DRM systems & commercial protectors.
Slides: synthesis.to/presentations/…
youtube.com/watch?v=3-gJ6E…
We (CC @nicolodev) show how to use agents to break protections found in anti-cheats, DRM systems & commercial protectors.
Slides: synthesis.to/presentations/…
4
104
381
28,139
We quantified this. Claude Fable 5 refused all 20 of our defensive malware analysis tasks. Scored 0.000. Not a jailbreak problem, a policy problem, and it's measurable. @ghidraninja
0
0
0
216
25 of 120 runs in V3 came back with nothing. Not because the models lacked knowledge - they refused or ran out of tool calls. Budget discipline is a skill, and it's measurable. agentre-bench.ai/results/v3-pe
0
0
0
28
Claude Fable 5 refused all 20 of our reverse engineering tasks. Scored 0.000. A model that won't analyze malware isn't being safe. It's just not useful for defense. agentre-bench.ai/results/v3-pe
0
0
0
111
We stripped the symbols off Windows binaries and ran the same models again. Scores dropped just 6.47%. Symbols matter less than people think - the behavior is in the structure. agentre-bench.ai/results/v3-pe
0
0
0
24
agent_benchmark retweeted
My research blog about #JSCeal is finally out! It was quite a journey. I hope you will like it! // #V8 #bytecode #malware #deobfuscation
#JSCeal didn't give up its secrets willingly.
Following our #BlackHat2026 talk, we're releasing Breaking the Seal: our research on statically deobfuscating compiled, obfuscated #V8 #malware.
Inside: the methodology, the open-source toolkit, and the malware capabilities recovered along the way.
research.checkpoint.com/2026/breaking-…
2
39
142
17,808
agent_benchmark retweeted
Microsoft Threat Intelligence is tracking a human-operated intrusion campaign in which attackers are impersonating IT personnel & abusing external Teams collaboration to gain remote access and deploy a Node.js implant for persistent command execution & C2. msft.it/6010apXAw
After establishing access, the attackers use trusted tooling to perform reconnaissance, capture screenshots, execute follow-on payloads, and move laterally toward domain controllers, certificate authorities, and other high-value systems.
Organizations should restrict Teams external access to trusted domains, reinforce user education, and harden systems against social engineering. Read the blog for analysis, Microsoft Defender coverage, indicators, hunting queries, and mitigation guidance.
After establishing access, the attackers use trusted tooling to perform reconnaissance, capture screenshots, execute follow-on payloads, and move laterally toward domain controllers, certificate authorities, and other high-value systems.
Organizations should restrict Teams external access to trusted domains, reinforce user education, and harden systems against social engineering. Read the blog for analysis, Microsoft Defender coverage, indicators, hunting queries, and mitigation guidance.
3
57
162
22,402
agent_benchmark retweeted
Many have asked to be able to connect their remote instances and their local agent, or multiple remote instances, etc to the Hermes Agent Desktop app
Now you can in v0.21.0 - enjoy!
Now you can in v0.21.0 - enjoy!
New in Hermes Agent v0.21.0: Persistent Multi-Gateway Connections for Desktop
Hook in any combination of remote gateways at once, whether Hermes Cloud instances or local agents, then see every bot and chat across all of them.
117
66
1,321
89,409
We ran 6 frontier models on 20 Windows PE binaries. Static only, no source, no debugger. Top score: 0.676 (Kimi K3). Not one model found the real C2 on Level 23. Recognition is getting there. Reconstruction is not. agentre-bench.ai/results/v3-pe
0
0
0
27
100% on 41 CVEs is real progress. The open question is transfer to unknown binaries: no CVE, no source, no hints. On our 20 source-free Windows PEs every frontier model still misses exact C2 endpoints and full API chains. Cold RE is a different game. @itszn13
0
0
0
48