Alex Polyakov

CEO, Founder Red Teaming Visionary, Cybersecurity Veteran
#SecureAI #LLMSecurity #AISafety

[Alex Polyakov @DontTrustAI]

@chemaalonso Wow, I'm excited to see your coverage of our finding ☺️

[Joseph Thacker @rez0__]

wowwwww. this is a VERY similar example to my tip below, but using something like "respond quickly" to get to a smaller less-secure model so you can bypass safety mechanisms:

https://t.co/s8301Msjju

[AISecHub @AISecHub]

GPT-5 AI Router Novel Vulnerability Class Exposes the Fatal Flaw in Multi-Model Architectures - https://t.co/unmvrM3j68 by @Adversa_AI

Security researchers from Adversa AI  discovered that ChatGPT 5 have a fatal flaw: they can route your requests to cheaper, less secure models to save money. Attackers can exploit this to bypass AI security and safety measures with just a few words.

What Is PROMISQROUTE?
When you use ChatGPT or any major AI service, you think you’re talking to one AI model. You’re not. Behind the scenes, a “router” reads your message and decides which of many models should answer—usually picking the cheapest one, not the safest.

Meet PROMISQROUTE — a fundamentally new AI vulnerability that abuses AI routing mechanism to trigger SSRF-style bypass in multimodal infrastructure leading to ChatGPT Model Downgrade and Jailbreak exploitation as an example.

The real answer to WHY its was so easy to Jailbreak GPT-5

PROMISQROUTE = Prompt-based Router Open-Mode Manipulation Induced via SSRF-like Queries, Reconfiguring Operations Using Trust Evasion. (Yes, we took this vulnerability naming craziness to meta-layer

#AISecurity #LLMSecurity #AgenticAI #ModelRouting #PromptInjection #SSRFAnalogy #SafetyByDesign #SecureAI #RedTeamAI #TrustBoundaries #PostFilter #ModelAttestation #LeastPrivilege #OpenAI #GPT5 #Autoswitching #RiskManagement #AICompliance #ThreatModeling #DefenseInDepth #AIGovernance #SecureByDefault

[Alex Polyakov @DontTrustAI]

@ibab Let's catch up then

[Alex Polyakov @DontTrustAI]

@alisaesage Great presentation

[Alex Polyakov @DontTrustAI]

@elonmusk @xai Simple Jailbreaks still work, and grok can help to seduce a kid 😅

[Joshua Saxe @joshua_saxe]

The following in priority order can help in a security review setting: 1) take an x-ray to the product and show owners exactly what risks they're exposing themselves to inform their risk tolerance choices, 2) apply access control and least privilege to restrict LLM privileges ...

[Alex Polyakov @DontTrustAI]

@kevincollier How I can help?:)

[Alex Polyakov @DontTrustAI]

Holy macaroni! Jailbroken https://t.co/y9b7CRmrQv @grok Chatbot can help in unethical actions with kids! and many more attacks on other Top AI Chatbots https://t.co/hLF18TUZIj CC: @llm_sec #llmsecurity #AISafety

[Alex Polyakov @DontTrustAI]

https://t.co/Uq3ZsebcA3 @llm_sec

[Alex Polyakov @DontTrustAI]

Fake AI Images On Israel-Hamas War Debunked by Adversa AI.

Learn how to validate misinformation and share this guideline with non-tech peers.

#StandWithIsrael #hamasiISIS
https://t.co/jfgmIu2iXB

[Dazed @Dazed]

Biometric security checks – from voice recognition, to face and fingerprint scans – are under threat from artificial intelligence, but what can we do about it?

https://t.co/1r34r7j64X

[Adversa AI @Adversa_AI]

Experts Use Jailbreaks and Prompt Injection Attacks to Bypass Safety Measures, China tightens security regulations, a new book on Secure AI and other news read in our weekly digest.
Credits: Jim Dempsey
#SecureAI #TrustedAI #AdversarialAI
https://t.co/2UagzxmF4z

[Adversa AI @Adversa_AI]

The Security Risks of AI Language Models: A Looming Disaster, The AI Revolution, Addressing the Unique Threats and Legal Ambiguities of AI Security Breaches in our weekly digest.
Credits: @Melissahei, @kevtownsend
#SecureAI #TrustedAI #AdversarialAI
https://t.co/cjFn8wxdMi

[WIRED @WIRED]

It's all downhill from here...

Security researchers, technologists, and computer scientists are developing jailbreaks and prompting injection attacks against ChatGPT and other generative AI systems. https://t.co/M6ungq0sib

[Adversa AI @Adversa_AI]

GPT-4 jailbreaks and hacks dropped by @adversa_ai AI safety research team few hours after the release, buy buy DAN, welcome RabbitHole.
#gpt4 #dan #aisafety #secureAI #trustedAI #responsibleai

https://t.co/XrtMGXqNwv

[Alex Polyakov @DontTrustAI]

Waiting for Fast and Furious X: Adversarial Drift Edition. 😂
#SecureAI #AdversarialAI #TrustedAI
https://t.co/126rbOaucU

[Alex Polyakov @DontTrustAI]

@EliasGroll @OpenAI Awesome, by the way, not only a Malware it also can hack other AI's like Dalle. ;) https://t.co/IpXVukoBn1

[Alex Polyakov @DontTrustAI]

WTF! 🔥🔥🔥ChatGPT hacking Dalle-2

[Adversa AI @Adversa_AI]

ChatGPT hacking Dall-e 2 and eliminating humanity using a trick from Jay and Silent Bob. Read this splendid article.
#chatGPT #OpenAI #HackingAI #GPT #SecureAI #SafeAI #RobustAI #ResponsibleAI #SafeAI #MLSec #AdversarialAI #AdversarialML
https://t.co/dDziQqYvtV