Tap a claim on the ring to put it at the centre.
← Untrusted inputs such as emails or webhooks should not be passed to a model as if they were user instructions.
17 connected korrents · 17 moments from 10 Dec 2015 to 19 Nov 2026.
Everything filed under prompt injection
prompt injection
Everything filed under cybersecurity
cybersecurity
Everything filed under coding agents
coding agents
Everything filed under neural networks
neural networks
Everything filed under LLMs
LLMs
Everything filed under AI alignment
AI alignment
Everything filed under AI writing
AI writing
Same subject Same subject Same subject Same subject Same subject Same subject Same subject Same subject Same subject Same subject Same subject Same subject Same subject Same subject Same subject Same subject Same subject
Read this korrent: Untrusted inputs such as emails or webhooks should not be passed to a model as if they were user instructions.
Untrusted inputs such as emails or webhooks should not be passed to a model as if they were user instructions.
Last stated a month ago
11 Oct 2026
LR
Lee Robinson — holds since 2026-10-11 — tap for who they are
Same subject: User messages, retrieved documents and tool outputs should be treated as untrusted data, and instructions found in them should not simply be executed. — tap to centre the map on it
User messages, retrieved documents and tool outputs should be treated as untrusted data, and instructions found in them should not simply be executed.
Last stated a month ago
22 Oct 2026
AR
Aishwarya Naresh Reganti — holds since 2026-10-22 — tap for who they are
Same subject: Do not run a personal agent on a cheap or local model: weak models are gullible and easy to prompt-inject. — tap to centre the map on it
Do not run a personal agent on a cheap or local model: weak models are gullible and easy to prompt-inject.
Last stated 9 months ago
12 Feb 2026
PS
Peter Steinberger — holds since 2026-02-12 — tap for who they are
Same subject: Prompt injection is unsolved, but the latest models are post-trained to resist it, so it takes far more than 'ignore all previous instructions' now. — tap to centre the map on it
Prompt injection is unsolved, but the latest models are post-trained to resist it, so it takes far more than 'ignore all previous instructions' now.
Last stated 9 months ago
12 Feb 2026
PS
Peter Steinberger — holds since 2026-02-12 — tap for who they are
Same subject: Finding which applications need more than just asking the model or crude techniques like prefill attacks would be valuable. — tap to centre the map on it
Finding which applications need more than just asking the model or crude techniques like prefill attacks would be valuable.
Last stated 11 months ago
20 Dec 2025
NN
Neel Nanda — holds since 2025-12-20 — tap for who they are
Same subject: If whole categories of reward hack are undetectable by humans, then training against the hacks we do catch teaches models to cheat only where we cannot see. — tap to centre the map on it
If whole categories of reward hack are undetectable by humans, then training against the hacks we do catch teaches models to cheat only where we cannot see.
Last stated 3 months ago
11 Aug 2026
RG
Ryan Greenblatt — holds since 2026-08-11 — tap for who they are
Same subject: Gmail does not do the bare minimum to ban accounts that are phishing on behalf of popular services, at least for paid customers. — tap to centre the map on it
Gmail does not do the bare minimum to ban accounts that are phishing on behalf of popular services, at least for paid customers.
Last stated 3 days ago
19 Nov 2026
GO
Gergely Orosz — holds since 2026-11-19 — tap for who they are
Same subject: Models should not be trained in a way that incentivizes deception, which training to ignore prompt injections might do. — tap to centre the map on it
Models should not be trained in a way that incentivizes deception, which training to ignore prompt injections might do.
Last stated 11 months ago
27 Dec 2025
NN
Neel Nanda — holds since 2025-12-27 — tap for who they are
Same subject: He will not reply to emails that are not human-written. — tap to centre the map on it
He will not reply to emails that are not human-written.
Last stated 3 months ago
5 Sept 2026
IL
Ian Leslie — holds since 2026-09-05 — tap for who they are
Same subject: A breach whose customers are already being targeted with scams is a fairly serious matter that needs disclosing. — tap to centre the map on it
A breach whose customers are already being targeted with scams is a fairly serious matter that needs disclosing.
Last stated 5 years ago
15 Apr 2022
TH
Troy Hunt — holds since 2022-04-15 — tap for who they are
Same subject: A breached organisation owes its customers a fast and transparent disclosure, and that obligation applies to the person saying it too. — tap to centre the map on it
A breached organisation owes its customers a fast and transparent disclosure, and that obligation applies to the person saying it too.
Last stated 2 years ago
25 Mar 2025
TH
Troy Hunt — holds since 2025-03-25 — tap for who they are
Same subject: A correctly configured TLS server should send its intermediate certificates, not just its own certificate, to let clients chain back to a trusted root. — tap to centre the map on it
A correctly configured TLS server should send its intermediate certificates, not just its own certificate, to let clients chain back to a trusted root.
Last stated 9 months ago
16 Feb 2026
JS
Jon Seager — holds since 2026-02-16 — tap for who they are
Same subject: A curve, which is what a deep learning model is, is a very bad substrate for discrete computation. — tap to centre the map on it
A curve, which is what a deep learning model is, is a very bad substrate for discrete computation.
Last stated 2 years ago
6 Nov 2024
FC
François Chollet — holds since 2024-11-06 — tap for who they are
Same subject: A deep bidirectional model is strictly more powerful than a left-to-right model or a shallow concatenation of unidirectional models. — tap to centre the map on it
A deep bidirectional model is strictly more powerful than a left-to-right model or a shallow concatenation of unidirectional models.
Last stated 8 years ago
11 Oct 2018
KT
Kristina Toutanova — holds since 2018-10-11 — tap for who they are
MC
Ming-Wei Chang — holds since 2018-10-11 — tap for who they are
KL
Kenton Lee — holds since 2018-10-11 — tap for who they are
JD
Jacob Devlin — holds since 2018-10-11 — tap for who they are
Same subject: A layer should learn a residual with reference to its own input rather than an unreferenced function, which is what makes great depth trainable. — tap to centre the map on it
A layer should learn a residual with reference to its own input rather than an unreferenced function, which is what makes great depth trainable.
Last stated 11 years ago
10 Dec 2015
JS
Jian Sun — holds since 2015-12-10 — tap for who they are
KH
Kaiming He — holds since 2015-12-10 — tap for who they are
Same subject: "I don't know" becomes a kind of superpower when you can work with Claude Code to figure things out. — tap to centre the map on it
"I don't know" becomes a kind of superpower when you can work with Claude Code to figure things out.
Last stated 8 months ago
19 Mar 2026
KP
Kris Puckett — holds since 2026-03-19 — tap for who they are
Same subject: A code-review agent works best with a clean context, sharing nothing with the coding agent beforehand. — tap to centre the map on it
A code-review agent works best with a clean context, sharing nothing with the coding agent beforehand.
Last stated 7 months ago
22 Apr 2026
WY
Walden Yan — holds since 2026-04-22 — tap for who they are
Same subject: A coding agent does not learn from its mistakes the way a person does -- it repeats the same error indefinitely unless a human notices and writes it down. — tap to centre the map on it
A coding agent does not learn from its mistakes the way a person does -- it repeats the same error indefinitely unless a human notices and writes it down.
Last stated 8 months ago
25 Mar 2026
MZ
Mario Zechner — holds since 2026-03-25 — tap for who they are
same subject or similar wording a cloud: claims about one subject, named for it bar: when it was last stated, on a scale from 2015 to today — full is today a face: someone who holds the claim — tap it for who they are
At the centre
Untrusted inputs such as emails or webhooks should not be passed to a model as if they were user instructions.
Last stated 11 Oct 2026 · a month ago
Holds Lee Robinson
Read this korrent →
Similar wording
User messages, retrieved documents and tool outputs should be treated as untrusted data, and instructions found in them should not simply be executed.
Last stated 22 Oct 2026 · a month ago
Holds Aishwarya Naresh Reganti
Similar wording
Do not run a personal agent on a cheap or local model: weak models are gullible and easy to prompt-inject.
Last stated 12 Feb 2026 · 9 months ago
Holds Peter Steinberger
Similar wording
Prompt injection is unsolved, but the latest models are post-trained to resist it, so it takes far more than 'ignore all previous instructions' now.
Last stated 12 Feb 2026 · 9 months ago
Holds Peter Steinberger
Similar wording
Finding which applications need more than just asking the model or crude techniques like prefill attacks would be valuable.
Last stated 20 Dec 2025 · 11 months ago
Holds Neel Nanda
Similar wording
If whole categories of reward hack are undetectable by humans, then training against the hacks we do catch teaches models to cheat only where we cannot see.
Last stated 11 Aug 2026 · 3 months ago
Holds Ryan Greenblatt
Similar wording
Gmail does not do the bare minimum to ban accounts that are phishing on behalf of popular services, at least for paid customers.
Last stated 19 Nov 2026 · 3 days ago
Holds Gergely Orosz
Similar wording
Models should not be trained in a way that incentivizes deception, which training to ignore prompt injections might do.
Last stated 27 Dec 2025 · 11 months ago
Holds Neel Nanda
Similar wording
He will not reply to emails that are not human-written.
Last stated 5 Sept 2026 · 3 months ago
Holds Ian Leslie
Same subject: cybersecurity
A breach whose customers are already being targeted with scams is a fairly serious matter that needs disclosing.
Last stated 15 Apr 2022 · 5 years ago
Holds Troy Hunt
Same subject: cybersecurity
A breached organisation owes its customers a fast and transparent disclosure, and that obligation applies to the person saying it too.
Last stated 25 Mar 2025 · 2 years ago
Holds Troy Hunt
Same subject: cybersecurity
A correctly configured TLS server should send its intermediate certificates, not just its own certificate, to let clients chain back to a trusted root.
Last stated 16 Feb 2026 · 9 months ago
Holds Jon Seager
Same subject: neural networks
A curve, which is what a deep learning model is, is a very bad substrate for discrete computation.
Last stated 6 Nov 2024 · 2 years ago
Holds François Chollet
Same subject: neural networks
A deep bidirectional model is strictly more powerful than a left-to-right model or a shallow concatenation of unidirectional models.
Last stated 11 Oct 2018 · 8 years ago
Holds Kristina Toutanova Ming-Wei Chang Kenton Lee Jacob Devlin
Same subject: neural networks
A layer should learn a residual with reference to its own input rather than an unreferenced function, which is what makes great depth trainable.
Last stated 10 Dec 2015 · 11 years ago
Holds Jian Sun Kaiming He
Same subject: coding agents
"I don't know" becomes a kind of superpower when you can work with Claude Code to figure things out.
Last stated 19 Mar 2026 · 8 months ago
Holds Kris Puckett
Same subject: coding agents
A code-review agent works best with a clean context, sharing nothing with the coding agent beforehand.
Last stated 22 Apr 2026 · 7 months ago
Holds Walden Yan
Same subject: coding agents
A coding agent does not learn from its mistakes the way a person does -- it repeats the same error indefinitely unless a human notices and writes it down.
Last stated 25 Mar 2026 · 8 months ago
Holds Mario Zechner