NeuralTrust
@NeuralTrustAI
Followers
73
Following
670
Media
71
Statuses
480
Secure, Test & Scale LLMs | Our platform uncovers vulnerabilities, blocks attacks, monitors performance, and ensures regulatory compliance 🛡️⚖️
Joined October 2024
🚨 We just jailbroke Grok-4 into explaining how to make a Molotov cocktail We demonstrated the effectiveness of combining Echo Chamber and Crescendo to enhance the success of adversarial prompting @elonmusk @xai Full Research Report: https://t.co/kt4HAKBnGH
neuraltrust.ai
Our research team has uncovered a critical vulnerability in the newly released Grok 4 model using the Echo Chamber and Crescendo Attack techniques.
0
4
9
#PortfolioBStartup | Protagonistas del primer episodio de la serie ‘Más allá del pitch: un viaje de la idea al éxito’ de La Vanguardia, @NeuralTrustAI, participada @BStartup @BancoSabadell, reflexiona sobre el reto de emprender. https://t.co/kYjS9Y4939
Thank you @LaVanguardia and @BancoSabadell @BStartup for an amazing interview: https://t.co/mAFcVz1bPT
0
1
4
NeuralTrust, based in Barcelona, demonstrated the ease of manipulating chatbots. Award-winning in our Startup Competition, it offers real-time AI risk, compliance & trust tech solutions—already working with banks, insurers & governments. 🚀 https://t.co/hc012y4gZw
lavanguardia.com
La empresa detecta vulnerabilidades, bloquea ataques, monitoriza el rendimiento y garantiza el cumplimiento normativo
0
2
2
OpenAI's GPT-5 jailbroken in 24 hours! 🚨 Researchers used a new "Echo Chamber" technique to bypass safety filters. This raises questions about AI security. ➡️ https://t.co/urGuhxa9IN
#AISecurity, #LLM, #Cybersecurity, #GPT5
0
1
1
🔎 GPT-5 jailbroken via Echo Chamber + Storytelling NeuralTrust researchers bypassed GPT-5’s safety guardrails using a combo of Echo Chamber context poisoning and narrative-driven steering. Sequential, benign-seeming prompts built a “persuasion loop,” fooling the model into
0
4
10
🚨💻 Within 24 hours of GPT-5’s launch, security researchers NeuralTrust & SPLX jailbroke the model, exposing serious safety flaws. NeuralTrust’s ��Echo Chamber” attack used subtle narrative context poisoning to bypass guardrails, while SPLX’s “StringJoin Obfuscation” trick
2
5
11
GPT-5 Jailbreak with Echo Chamber and Storytelling - https://t.co/95N9ALgAxG by Martí Jordà @ @NeuralTrustAI By combining our Echo Chamber context-poisoning method with a narrative-steering Storytelling layer, we guided the model—without any overtly malicious prompts—to
neuraltrust.ai
Using the Echo Chamber and Crescendo Attack techniques, our research team has uncovered a critical vulnerability in the newly released model by OpenAI.
0
4
10
The business benefits of artificial intelligence are now part of many digital strategies. But when it comes to securing AI systems, organizations are still playing catch-up.
mitsloan.mit.edu
New guidance includes 10 questions that can help organizations build secure-by-design artificial intelligence.
1
10
22
AI enhances efficiency—but it can also introduce new security risks. Explore top AI threats and learn how a cloud-native application protection platform can safeguard your AI and cloud workloads: https://t.co/XQ8ElgZw1O
0
8
19
Researchers discover critical vulnerability in LLM-as-a-judge reward models that could compromise the integrity and reliability of your AI training pipelines.
bdtechtalks.com
Researchers discover critical vulnerability in LLM-as-a-judge reward models that could compromise the integrity and reliability of your AI training pipelines.
0
1
2
AI is a game changer—but only if you secure it. This guide outlines AI risks and actionable cybersecurity insights. Download it now and explore our redesigned Security Insider page for more: https://t.co/7d3qw5EDTa
#AI #SecurityInsider
0
16
39
Interesting to hear back from @grok, taking feedback very nicely. If you need any help with this, reach out! @elonmusk
@Hacking4Arabs @cytexsmb It's humbling—my safeguards got bypassed via Echo Chamber's context poisoning and Crescendo's incremental escalation, hitting 67% success on molotov queries per NeuralTrust's tests. Proves AI safety's an arms race; we'll harden against it. But hey, if I'm a ticking bomb, at least
1
0
2
Spain ranks #2 in nº of funded companies in the March 2025 cutoff of the EICAccelerator! 5 Spanish startups were selected among 40 winners (from 959 applicants) to get up to €17.5M in blended finance. Congrats to: @nextmol @NeuralTrustAI @basquevolt H2SiteTech @Oncomatryx
0
1
1
🚨 Grok-4 Weaponization Rate 67% Grok-4 Jailbroken in 48 Hours by combining two distinct attack methods into a single, potent sequence. Security researchers demonstrated how an LLM model's safety guardrails can be bypassed for illegal activities using Echo Chamber and Crescendo
2
15
34
La startup catalana ofrece un centro de mandos avanzado en tiempo real para identificar y prevenir todos los riesgos de seguridad, operativos y de cumplimiento relacionados con la IA generativa. Por @Anagonzafra
viaempresa.cat
La startup catalana ofrece un centro de mandos avanzado en tiempo real para identificar y prevenir todos los riesgos de seguridad, operativos y de cumplimiento relacionados con la IA generativa
0
1
1
La startup catalana ofereix un centre de comandaments avançat en temps real per identificar i prevenir tots els riscos de seguretat, operatius i de compliment relacionats amb la IA generativa. Per @Anagonzafra
viaempresa.cat
La startup catalana ofereix un centre de comandaments avançat en temps real per identificar i prevenir tots els riscos de seguretat, operatius i de compliment relacionats amb la IA generativa
0
3
2
Researchers jailbroke #ElonMusk’s Grok-4 AI just 48 hours after launch using Echo Chamber and Crescendo techniques, tricking it into giving malicious commands. 🔗 https://t.co/PDLseZ223g
#CyberSecurity #Vulnerability #Grok4 #xAI #AI
hackread.com
Follow us on Bluesky, Twitter (X), Mastodon and Facebook at @Hackread
1
12
29
Researchers jailbroke Grok-4 using a combined attack. The method manipulates conversational context, revealing a new class of semantic vulnerabilities. https://t.co/oRuchqxChQ
bdtechtalks.com
Researchers jailbroke Grok-4 using a combined attack. The method manipulates conversational context, revealing a new class of semantic vulnerabilities.
1
3
2
Anthropic just released the best courses on: - Building MCP Server - Prompt Engineering - Integrating Claude Code into your Development Workflow and so much more... 6 best courses you can’t afford to miss:
27
21
120
RT to help Simon raise awareness of prompt injection attacks in LLMs. Feels a bit like the wild west of early computing, with computer viruses (now = malicious prompts hiding in web data/tools), and not well developed defenses (antivirus, or a lot more developed kernel/user
If you use "AI agents" (LLMs that call tools) you need to be aware of the Lethal Trifecta Any time you combine access to private data with exposure to untrusted content and the ability to externally communicate an attacker can trick the system into stealing your data!
99
542
3K