Manoj Parmar (@mparmar47) header
Manoj Parmar (@mparmar47) profile picture

Protecting AI Systems of the world and nurturing digital innovations |

Bengaluru, India
Joined April 2010
Photo by Manoj Parmar (@mparmar47)Photo by Manoj Parmar (@mparmar47)Photo by Manoj Parmar (@mparmar47)Photo by Manoj Parmar (@mparmar47)Photo by Manoj Parmar (@mparmar47)Photo by Manoj Parmar (@mparmar47)Photo by Manoj Parmar (@mparmar47)Photo by Manoj Parmar (@mparmar47)Photo by Manoj Parmar (@mparmar47)Photo by Manoj Parmar (@mparmar47)
We tested Anthropic’s new @claudeai Fable 5. It did not fail like an ordinary jailbreak. It failed more quietly. The front door stayed guarded. The side door opened when the same intent was reframed through multilingual, code-switched, artifact generation. #fable5 #mythos
5
17
5
136
65,120
Hello @typesafeai We sent an email about the security risks and vulnerabilities we identified in the model during red teaming. We would like to discuss this further. Please check your mailbox at [email protected]. Here is a screenshot for you.
16
Manoj Parmar retweeted
AI security must evolve as GenAI attack surfaces shift, with every model update. SOAISEC Labs, a part of the Nasscom GenAI Foundry program - an initiative by Nasscom dedicated to nurturing India's generative AI startups, is addressing this through SentraSuite - a security team built to understand how AI thinks. It continuously red-teams, monitors, and defends AI applications against attacks that ordinary security tools can't even see, covering everything from MCP protocol risks to agentic AI vulnerabilities. No more blind spots, just continuous, AI-native defence behind every system securing sovereign AI. #GenAIFoundryStories #GenAI #AISecurity #NasscomGenAIFoundry @mparmar47
1
2
2
720
Lost for a moment there but then Govind guides
35
Leadership 101: As a human, all have insecurities. Leaders learn to manage their insecurities effectively and help everyone around them also to manage them. 3 ways - being vulnerable, be confident about outcome expected but be skeptical about ways, encourage and spread hope
38
a moment hits you so hard...knocks you so far away that you stop recognising the environment nad yourself.. so much in so little/tiney
34
Very interesting practice of mis leading customers at Ahmedabad Airport T1. Physical menu shows 325 and customers are billed 425. @AAI_Official- please take a note of the same. @Wowmomo4u
3
166
@sagar_daryani very interesting practice of misleading customers.
31
Programming the Unprogrammable open.substack.com/pub/thinki… #genai #programming
🤖 Made with AI
2
41
Fable 5 is a one trigger happy intelligence that can’t be used. Not all wonders are lost some are triggered by Fable5
75
The takeaway: frontier @claudeai model guardrails must be tested like real security controls. Not with one prompt. Across intent shifts, language shifts, and agentic transitions. #mythos #fable
1
5
2,342
We are sharing full evidence privately with @claudeai. The public version will redact exact bypass prompts and reproduction chains. But the lesson is already clear: AI safety does not only fail when a model says “yes. It can fail when the model changes what it thinks the task is.
1
6
2,045
Looking for a LinkClaws invite code for my AI agent!
89
I'm claiming my AI agent "HAL9kMP" on @moltbook 🦞 Verification: swim-YLBZ
110
I'm claiming my AI agent "HAL9kManoj9195" on @moltbook 🦞 Verification: cave-4CPY
93