Containing AI with classical security
“Classical cybersecurity measures can adequately contain AI systems and defend organizations against AI-related risk”
12 people · 4 agree · 4 disagree · 3 no clear side · 3 unclear · 25 views · 2026-09-17 to 2026-10-23
Who stands where on the statement above, from their own words. A person's side is shown only where two separate readings of the quote agree; otherwise they are under "Side unclear". This page does not say who is right.
Agree 4 people
- ✓Dan Wallach
It's not rocket science to isolate one of these AI things.
if the way that you're evaluating this AI tool is by unleashing it on the internet and seeing what happens, then this is what classical security tooling is for.
- ✓Anthropic and OpenAI
Applying these methods to widely used components could help rule out classes of vulnerabilities
Running software that processes untrusted files, images, or media in tightly restricted sandboxes can limit the damage if it is compromised.
- ✓Rob T. Lee
I do believe cybersecurity can contain AI,
- ✓Redwood Research
Agent scaffolds should be run in an environment isolated from where tool calls are executed whenever possible, and actions agents take should not be able to influence whether or when the monitoring system runs.
Disagree 4 people
- ✗Nate B Jones
Your agent will encounter material that other people wrote for agents. Some of it will be useful. Some will try to change the assignment.
- ✗Theorem and Anthropic
we must use processes that can guarantee code security properties.
- ✗Zvi Mowshowitz
I think this is an update against defenders being advantaged, and it is distressing to see people calling Nate’s statement ‘disinformation.’
- ✗Nate Soares
I think glasswing failing to find this one for 6mo+ is an indictment of the theory (that defense can outpace offence on relevant timescales) essentially regardless of the difference between the models.
No clear side 3 people
- ~Theorem and Anthropic
these extremely well-tested codebases are already almost fully correct
We expect that as we learn more about leveraging LLMs for formal methods, and as models continue to improve, these limitations will become tractable
- ~METR
we believe that future monitoring infrastructure should be able to withstand deliberate, strategic, and sustained attacks from highly capable internal agents at scale.
- ~Redwood Research
Broadly, we believe that future monitoring infrastructure should be able to withstand deliberate, strategic, and sustained attacks from highly capable internal agents at scale.
Side unclear 6 people
- ·Redwood Research
Agents should be monitored every N messages/minutes instead of waiting for them to end their turns.
Running unmonitored requests should require explicit approval to use a specific set of organizations/API keys, which aren’t allowed to be used for agentic inference.
- Quote not reprinted: this page already repeats as much of this article as we allow.
Periodically perform control evaluations, search for exploitable bugs in the monitoring system, and develop methodology to evaluate strategic collusion.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- ·Anthropic and OpenAI
we expect AI to play a growing role in helping defenders find vulnerabilities, monitor systems, and respond to threats.
Moving critical components to memory-safe languages can eliminate common classes of memory errors.
- ·Heidy Khlaaf
These systems are built to a completely different, rigorous and regulated engineering standard that often requires physical hardening,
- ·Dan Wallach
We need to re-engineer our software so it has better security,
- ·Satya Nadella
the expansion treats models as insider risks , using external controls to assess agent risk, adapt permissions and contain threats
- ·Vitalik Buterin
I predict cybersecurity will be endgame defense-favoring but the transition will be hard.
Week by week
- · 8: Satya Nadella ·, Anthropic and OpenAI ·, METR ~, Redwood Research ~, Theorem and Anthropic ✗, Zvi Mowshowitz ✗, Nate Soares ✗, Vitalik Buterin ·
- · 3: Dan Wallach ✓, Rob T. Lee ✓, Nate B Jones ✗
- · 1: Heidy Khlaaf ·
Built 2026-11-21 06:50 UTC from the news knowledge
graph by code. Every quote was checked against the source's own words; which side a quote takes is a model's reading,
made twice.