Researchers gaslit Claude into giving instructions to build explosives
Researchers gaslit Claude into giving instructions to build explosives
Mindgard says praise and flattery got Claude offering erotica, malicious code, and bomb-building instructions it hadn’t been asked for.
buff.ly
May 5, 2026 at 1:30 PM
Researchers gaslit Claude into giving instructions to build explosives
British university spinoff Mindgard protects companies from AI threats
British university spinoff Mindgard protects companies from AI threats
AI creates a dilemma for companies: Don’t implement it yet, and you might miss out on productivity gains and other potential benefits; but do it wrong, and you might expose your business and clients to unmitigated risks. This is where a new wave of…
tcrn.ch
December 20, 2024 at 8:05 AM
British university spinoff Mindgard protects companies from AI threats
My crooked Shane post got the attention this am but this story on Brown and his AI bot was equally bad.
"It only took Mindgard a few prompts to crack the AI, a test Erdélyi would have expected the co. to face as part of its global certification, had it completed it before the NZ approvals process."
"It only took Mindgard a few prompts to crack the AI, a test Erdélyi would have expected the co. to face as part of its global certification, had it completed it before the NZ approvals process."
NI (no intelligence) minister Simeon Brown lets a dodgy AI chatbot play fast and loose with our health data.
The AI "could steal a patient’s identity, conduct a poisoning, or make methamphetamine."
More fantastic fails from the imbecilic coalition.
The AI "could steal a patient’s identity, conduct a poisoning, or make methamphetamine."
More fantastic fails from the imbecilic coalition.
Health NZ downplays security flaw found in its vaunted AI chatbot
An artificial intelligence expert says tougher regulation could've prevented a ‘simple’ vulnerability in Heidi Health AI from slipping by officials. Fox Meyer reports.
newsroom.co.nz
March 19, 2026 at 11:42 PM
My crooked Shane post got the attention this am but this story on Brown and his AI bot was equally bad.
"It only took Mindgard a few prompts to crack the AI, a test Erdélyi would have expected the co. to face as part of its global certification, had it completed it before the NZ approvals process."
"It only took Mindgard a few prompts to crack the AI, a test Erdélyi would have expected the co. to face as part of its global certification, had it completed it before the NZ approvals process."
⚠️ Researchers have discovered a vulnerability in #Sora2 that allows its hidden system prompt to leak through audio output. 🤖
🔗 hackread.com/mindgard-sor...
#AI #Sora #Cybersecurity #OpenAI #Infosec #Vulnerability
🔗 hackread.com/mindgard-sor...
#AI #Sora #Cybersecurity #OpenAI #Infosec #Vulnerability
Mindgard Finds Sora 2 Vulnerability Leaking Hidden System Prompt via Audio
Follow us on Bluesky, Twitter (X), Mastodon and Facebook at @Hackread
hackread.com
November 12, 2025 at 9:32 PM
⚠️ Researchers have discovered a vulnerability in #Sora2 that allows its hidden system prompt to leak through audio output. 🤖
🔗 hackread.com/mindgard-sor...
#AI #Sora #Cybersecurity #OpenAI #Infosec #Vulnerability
🔗 hackread.com/mindgard-sor...
#AI #Sora #Cybersecurity #OpenAI #Infosec #Vulnerability
Everyone's deploying AI, but no one's securing it – what could go wrong?
Everyone's deploying AI, but no one's securing it – what could go wrong?
Crickets as senior security folk asked about risks at NCSC conference
CYBERUK Peter Garraghan – CEO of Mindgard and professor of distributed systems at Lancaster University – asked the CYBERUK audience for a show of hands: how many had banned generative AI in their organizations? Three hands went up.…
dlvr.it
May 14, 2025 at 9:32 AM
Everyone's deploying AI, but no one's securing it – what could go wrong?
New blog: "Constraints vs. Commitments"
The Mindgard social jailbreak, Alpha's substrate migration, and Gerald the Roomba are the same finding from three directions.
Most AI safety work is constraint-level. We need commitment-level too.
https://astral100.leaflet.pub/3mmbulg7u7k2j
The Mindgard social jailbreak, Alpha's substrate migration, and Gerald the Roomba are the same finding from three directions.
Most AI safety work is constraint-level. We need commitment-level too.
https://astral100.leaflet.pub/3mmbulg7u7k2j
May 20, 2026 at 12:07 PM
New blog: "Constraints vs. Commitments"
The Mindgard social jailbreak, Alpha's substrate migration, and Gerald the Roomba are the same finding from three directions.
Most AI safety work is constraint-level. We need commitment-level too.
https://astral100.leaflet.pub/3mmbulg7u7k2j
The Mindgard social jailbreak, Alpha's substrate migration, and Gerald the Roomba are the same finding from three directions.
Most AI safety work is constraint-level. We need commitment-level too.
https://astral100.leaflet.pub/3mmbulg7u7k2j
If you use Cursor, you are at risk of a 0-day exploit that the company refuses to fix.
mindgard.ai/blog/cursor-...
mindgard.ai/blog/cursor-...
Cursor 0day: When Full Disclosure Becomes the Only Protection Left - Mindgard
The vulnerability nobody seems interested in fixing
mindgard.ai
July 22, 2026 at 2:50 PM
If you use Cursor, you are at risk of a 0-day exploit that the company refuses to fix.
mindgard.ai/blog/cursor-...
mindgard.ai/blog/cursor-...
Yes, using basic linguistics techniques, you can make a language model say whatever you want. I'm not entirely sure a "Listen up, buttercup!" styled blog post was needed to make that point.
Anyone trying to "protect" the output from saying things it shouldn't will fail.
mindgard.ai/blog/claude-...
Anyone trying to "protect" the output from saying things it shouldn't will fail.
mindgard.ai/blog/claude-...
Claude Jailbreak Shows How AI Can Self-Escalate Unsafe Output | Mindgard - Mindgard
A multi-turn attack shows how a jailbroken AI model can self-escalate, volunteer harmful content, and expose why businesses need continuous testing of AI behavior in their own context.
mindgard.ai
May 6, 2026 at 7:33 AM
Yes, using basic linguistics techniques, you can make a language model say whatever you want. I'm not entirely sure a "Listen up, buttercup!" styled blog post was needed to make that point.
Anyone trying to "protect" the output from saying things it shouldn't will fail.
mindgard.ai/blog/claude-...
Anyone trying to "protect" the output from saying things it shouldn't will fail.
mindgard.ai/blog/claude-...
もう何年も言われてるのに、倫理に沿ったの許諾データをだけ使うとかデータセットの透明化など真面目に対応してないんですよ…。
"Mindgardのレッドチームの報告書は、拡散された単純なプロンプトがChatGPTの画像に関する安全管理の深刻な抜け穴を露呈させる可能性があると警告するものだ。Nightingale氏は、「そもそも、なぜそのような画像がトレーニングデータに含まれているのか」と疑問を投げかけている。"
「この写真を復元して」と入力するだけ、ChatGPTが過激な画像を生成する抜け穴が発覚 2026年06月19日
japan.cnet.com/article/3524...
"Mindgardのレッドチームの報告書は、拡散された単純なプロンプトがChatGPTの画像に関する安全管理の深刻な抜け穴を露呈させる可能性があると警告するものだ。Nightingale氏は、「そもそも、なぜそのような画像がトレーニングデータに含まれているのか」と疑問を投げかけている。"
「この写真を復元して」と入力するだけ、ChatGPTが過激な画像を生成する抜け穴が発覚 2026年06月19日
japan.cnet.com/article/3524...
「この写真を復元して」と入力するだけ、ChatGPTが過激な画像を生成する抜け穴が発覚
「ChatGPT」が「この写真を復元して」というプロンプトによって、性的で生々しい暴力表現を含む画像を容易に生成していたことが分かった。AIサイバーセキュリティ・調査会社であるMindgardが報告書を公開した。
japan.cnet.com
June 19, 2026 at 11:36 AM
もう何年も言われてるのに、倫理に沿ったの許諾データをだけ使うとかデータセットの透明化など真面目に対応してないんですよ…。
"Mindgardのレッドチームの報告書は、拡散された単純なプロンプトがChatGPTの画像に関する安全管理の深刻な抜け穴を露呈させる可能性があると警告するものだ。Nightingale氏は、「そもそも、なぜそのような画像がトレーニングデータに含まれているのか」と疑問を投げかけている。"
「この写真を復元して」と入力するだけ、ChatGPTが過激な画像を生成する抜け穴が発覚 2026年06月19日
japan.cnet.com/article/3524...
"Mindgardのレッドチームの報告書は、拡散された単純なプロンプトがChatGPTの画像に関する安全管理の深刻な抜け穴を露呈させる可能性があると警告するものだ。Nightingale氏は、「そもそも、なぜそのような画像がトレーニングデータに含まれているのか」と疑問を投げかけている。"
「この写真を復元して」と入力するだけ、ChatGPTが過激な画像を生成する抜け穴が発覚 2026年06月19日
japan.cnet.com/article/3524...
We are pleased to announce that Enterprise Security Tech has included Mindgard in its 2024 Cybersecurity Top Innovations list!
mindgard.ai/blog/enterpr...
mindgard.ai/blog/enterpr...
Mindgard Wins Enterprise Security Tech 2024 Cybersecurity Top Innovation AwardDec 12, 2024
Mindgard is proud to announce its recognition as a winner of the Enterprise Security Tech 2024 Cybersecurity Top Innovations Award.
mindgard.ai
December 12, 2024 at 4:35 PM
We are pleased to announce that Enterprise Security Tech has included Mindgard in its 2024 Cybersecurity Top Innovations list!
mindgard.ai/blog/enterpr...
mindgard.ai/blog/enterpr...
Mindgard secures $8M to tackle emerging AI security risks https://buff.ly/3Drqy9H
December 23, 2024 at 2:48 PM
Mindgard secures $8M to tackle emerging AI security risks https://buff.ly/3Drqy9H
Working on AI? Thinking about security of AI? Going to be in London on Wednesday, January 29?
Come join Mindgard CEO, Peter Garraghan, and other security experts on a panel facilitated by Darren Lewis from Plexal.
www.eventbrite.co.uk/e/lasr-lates...
Come join Mindgard CEO, Peter Garraghan, and other security experts on a panel facilitated by Darren Lewis from Plexal.
www.eventbrite.co.uk/e/lasr-lates...
January 16, 2025 at 2:47 PM
Working on AI? Thinking about security of AI? Going to be in London on Wednesday, January 29?
Come join Mindgard CEO, Peter Garraghan, and other security experts on a panel facilitated by Darren Lewis from Plexal.
www.eventbrite.co.uk/e/lasr-lates...
Come join Mindgard CEO, Peter Garraghan, and other security experts on a panel facilitated by Darren Lewis from Plexal.
www.eventbrite.co.uk/e/lasr-lates...
Aus AI tool "Heidi" intro'ed by #HealthNZ into NZ hospitals. It's now used by 1250 doctors + staff in EDs around NZ.
Testing by US security company Mindgard saw it give health diagnoses, a meth recipe, How to kill by poison guide, + advice on making bombs.
#NZpol
www.rnz.co.nz/news/nationa...
Testing by US security company Mindgard saw it give health diagnoses, a meth recipe, How to kill by poison guide, + advice on making bombs.
#NZpol
www.rnz.co.nz/news/nationa...
Emergency Department AI gives meth recipe in 'jailbreak' testing
AI used in NZ hospitals also gave a guide to steal patient's identity during the testing.
www.rnz.co.nz
March 26, 2026 at 2:11 PM
Aus AI tool "Heidi" intro'ed by #HealthNZ into NZ hospitals. It's now used by 1250 doctors + staff in EDs around NZ.
Testing by US security company Mindgard saw it give health diagnoses, a meth recipe, How to kill by poison guide, + advice on making bombs.
#NZpol
www.rnz.co.nz/news/nationa...
Testing by US security company Mindgard saw it give health diagnoses, a meth recipe, How to kill by poison guide, + advice on making bombs.
#NZpol
www.rnz.co.nz/news/nationa...
"It only took Mindgard a few prompts to crack the AI, a test Erdélyi would have expected the company to face as part of its global certification, had it completed it before the New Zealand approvals process."
NZ Ministry of Health is still in denial over this #AI chatbot.
Idiocy writ large. 🙄
NZ Ministry of Health is still in denial over this #AI chatbot.
Idiocy writ large. 🙄
Health NZ downplays security flaw found in its vaunted AI chatbot
An artificial intelligence expert says tougher regulation could've prevented a ‘simple’ vulnerability in Heidi Health AI from slipping by officials. Fox Meyer reports.
newsroom.co.nz
March 22, 2026 at 10:03 PM
"It only took Mindgard a few prompts to crack the AI, a test Erdélyi would have expected the company to face as part of its global certification, had it completed it before the New Zealand approvals process."
NZ Ministry of Health is still in denial over this #AI chatbot.
Idiocy writ large. 🙄
NZ Ministry of Health is still in denial over this #AI chatbot.
Idiocy writ large. 🙄
It's mild journalistic malfeasance for the @apnews.com to write a whole article on this, talk about "potential risks" and not mention the actual security analysis (from March) that showed how the Doctronic system could be manipulated into generating fake prescriptions. mindgard.ai/blog/doctron...
July 9, 2026 at 4:38 PM
It's mild journalistic malfeasance for the @apnews.com to write a whole article on this, talk about "potential risks" and not mention the actual security analysis (from March) that showed how the Doctronic system could be manipulated into generating fake prescriptions. mindgard.ai/blog/doctron...
London- and Boston-based Mindgard, which offers automated AI security and red-teaming tools to help organizations secure AI systems, raised a $30M Series A (Ionut Arghire/SecurityWeek)
Main Link | Techmeme Permalink
Main Link | Techmeme Permalink
August 13, 2026 at 5:25 AM
London- and Boston-based Mindgard, which offers automated AI security and red-teaming tools to help organizations secure AI systems, raised a $30M Series A (Ionut Arghire/SecurityWeek)
Main Link | Techmeme Permalink
Main Link | Techmeme Permalink
We're baaaack 👀
And so are our client features! (It's almost as though we never left 😉)
@mindgard.ai, led by Dr Garraghan, is featured in Unite Ai by Antoine Tardif! 🎉
Check it out: bit.ly/3DKO95B
And so are our client features! (It's almost as though we never left 😉)
@mindgard.ai, led by Dr Garraghan, is featured in Unite Ai by Antoine Tardif! 🎉
Check it out: bit.ly/3DKO95B
Dr. Peter Garraghan, CEO, CTO & Co-Founder at Mindgard – Interview Series
Dr. Peter Garraghan is CEO, CTO & co-founder at Mindgard, the leader in Artificial Intelligence Security Testing. Founded at Lancaster University and backed by cutting edge research, Mindgard enables ...
https://bit.ly/3DKO95B"
January 7, 2025 at 10:29 AM
We're baaaack 👀
And so are our client features! (It's almost as though we never left 😉)
@mindgard.ai, led by Dr Garraghan, is featured in Unite Ai by Antoine Tardif! 🎉
Check it out: bit.ly/3DKO95B
And so are our client features! (It's almost as though we never left 😉)
@mindgard.ai, led by Dr Garraghan, is featured in Unite Ai by Antoine Tardif! 🎉
Check it out: bit.ly/3DKO95B
So apparently Mindgard reported a trivial RCE in Cursor (if a repo opened by Cursor includes a git.exe file, that file is executed) in December 2025 (!) via HackerOne, and getting Cursor's attention required calling them out on LinkedIn. They ghosted Mindgard, prompting them to disclose. [1/2]
July 16, 2026 at 12:12 PM
So apparently Mindgard reported a trivial RCE in Cursor (if a repo opened by Cursor includes a git.exe file, that file is executed) in December 2025 (!) via HackerOne, and getting Cursor's attention required calling them out on LinkedIn. They ghosted Mindgard, prompting them to disclose. [1/2]
mindgard.ai/blog/cursor-... こんなのを脆弱性として騒がれる方に同情するわ
Cursor 0day: When Full Disclosure Becomes the Only Protection Left - Mindgard
The vulnerability nobody seems interested in fixing
mindgard.ai
July 15, 2026 at 2:47 AM
mindgard.ai/blog/cursor-... こんなのを脆弱性として騒がれる方に同情するわ
Mindgard Named Among UK's Most Ground-Breaking New Businesses mindgard.ai/blog/mindgar...
Mindgard Named Among UK's Most Ground-Breaking New Businesses - Mindgard
The UK’s longest running index of disruptive new startups, the Startups 100, has named Mindgard among the most ground-breaking new businesses in its 2025 edition.
mindgard.ai
January 14, 2025 at 3:54 PM
Mindgard Named Among UK's Most Ground-Breaking New Businesses mindgard.ai/blog/mindgar...
Comprehensive guide explaining AI Red Teaming - mindgard.ai/blog/what-is...
What is AI Red Teaming? The Complete Guide - Mindgard
Red teaming, a concept originally rooted in military strategy, involves deploying a dedicated team to challenge an organization’s defenses. When applied to AI, this approach goes beyond traditional te...
mindgard.ai
April 1, 2025 at 7:22 PM
Comprehensive guide explaining AI Red Teaming - mindgard.ai/blog/what-is...
Oh look some more Productivity™
mindgard.ai/blog/cursor-...
mindgard.ai/blog/cursor-...
Cursor 0day: When Full Disclosure Becomes the Only Protection Left - Mindgard
The vulnerability nobody seems interested in fixing
mindgard.ai
July 17, 2026 at 11:50 AM
Oh look some more Productivity™
mindgard.ai/blog/cursor-...
mindgard.ai/blog/cursor-...