Anthropic
An organization in the news record
| Kind | An organization · graph entity E14 · Wikidata Q116758847 |
|---|---|
| Wikipedia's sources | 1 of the 12 references dated 2026-09-01 to 2026-11-14 in Anthropic · lacking: reuters.com 4, techcrunch.com 2, theguardian.com 1, bloomberg.com 1, cnbc.com 1, sfstandard.com 1 and 1 more · 39 more from the year before the record |
| Views | 113 quoted, from 15 source domains |
| Dates | 2026-06-05 to 2026-11-12 · 11 weeks |
| Arguments | 22 · cyber risk from frontier AI, Human code review role, AI and jobs/workers, Measurement-driven AI optimization, AI and quantum computing timeline, Anthropic business risks … |
| Reported by | Above the Law, Anthropic, Ars Technica, Fortune, Inside Higher Ed, Interconnects, Latent Space, METR and 5 more |
| Profile | a readable profile, written from these views and checked sentence by sentence |
| Also | the lab page (models, prices, deals) |
Only what the organization says as itself. Views of its staff, teams or founders, named or not, are not counted here.
Where they stand
By argument, from their own words. ✓ agree · ✗ disagree · ~ no clear side · · side unclear (the two readings differ).
AI incidents and loss of control ✗
- ✗
the safeguards deployed on our generally available models would have blocked the behaviors identified
- ✗
These facts give us cautious optimism that with tighter monitoring and controls around evaluation infrastructure, as well as continued investment in alignment, this type of risk can be overcome.
cyber risk from frontier AI ✓ ~
- ✓
Our experiments suggest that small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things (IoT) devices
- ✓
If small models improve at their historical rate, they’ll match this capability level by the end of this year.
- ✓
A replicator like this would be challenging to shut down, since there’s no centralized point to shut off—though it may be slow (with a doubling time of days).
- ~
Our evaluation keeps the cyber-exploit element of the replication trivial (there’s no authentication over the ADB port), which makes this an upper bound on real-world success.
- ✓
An AI botnet isn’t tied to any particular hardware or software type, as evidenced by small models being able to set up inference engines on machines from a Jetson Nano to a Mac
- ~ Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ Quote not reprinted: this page already repeats as much of this article as we allow.
- ~
We’d expect quarantining to be similarly effective for an initial defense against a weight-based replicator.
- ✓ Quote not reprinted: this page already repeats as much of this article as we allow.
Human code review role ✓ ~ ·
- ✓
We think "did a human read the code?" is an imprecise proxy for "what could go wrong with this change, and which check—automated or human—is the right one to catch it?"
- ✓
When you are shipping this fast, manual, line-by-line code review is incredibly difficult if not impossible.
- ~
Review has to converge. An agent author and an agent reviewer can keep each other busy indefinitely.
- ✓
If the goal is to spend human attention well, this beats reading trivial diffs line by line.
- ·
Never let a bot's approval replace a required human's; dismiss approvals on new commits.
- ~
Noise costs more when the author is an agent. A person can wave off a nit in seconds.
- ✓
Write down risk tiers, set from the change rather than the author; start strict and loosen with evidence.
- ✓ Quote not reprinted: this page already repeats as much of this article as we allow.
AI and jobs/workers ✓ ·
- ✓
Whatever the new deal between AI-affected employees and their employers, a mechanism to share benefits of adoption is essential.
- ✓
Building systems to flexibly reallocate workers internally will be critical for AI-intensive organizations.
- ✓
Private action on the part of employers must come alongside farsighted government policy around retraining and income support.
- ✓
Reading these patterns forward and out across the economy suggests that AI diffusion could risk worker displacement.
- ✓
prioritizing reallocation forces AI-adopting companies to consider ways that AI can create new tasks for new products and new markets, rather than funneling AI applications into replacing human workers to save on labor costs.
- ✓
None of these actions will prevent powerful AI from disrupting jobs. And indeed none directly address the slowdown in junior hires documented in this report.
- ·
Uncertainty about the future set of these tasks should recommend preserving a flexible workforce.
- ✓ Quote not reprinted: this page already repeats as much of this article as we allow.
Measurement-driven AI optimization ·
- ·
With Claude, measuring something makes it tractable. Measurement used to be step zero: you’d add a metric, wait for data to roll in, and only then start to understand the problem.
- ·
This meant the highest-leverage thing we could do was find more things to measure.
- ·
We also knew that performance wins decay in a fast-moving codebase, and code ships fast at Anthropic. Once a project proved a win, we invested in ways to protect it.
- ·
We expect to keep working this way, one thread at a time, at any scale.
- ·
anything we could count, Claude could climb
AI and quantum computing timeline ·
- ·
Such improvements substantially reduce the quantum resources that useful applications require, so practical quantum computers could arrive sooner than previously anticipated.
- ·
If AI can speed up that work, useful quantum computing could arrive sooner.
- ·
AI could optimize all of them jointly, from the algorithm down to the hardware, to build an entire error-corrected quantum computer tailored to a specific problem.
- ·
These reductions largely came from redesigning the circuits that carry out the computation, rather than from making better hardware.
Anthropic business risks ·
- ·
are not subject to long-term contracts with us,” Anthropic writes. That means they could lower their spending
- ·
creates complex dynamics that could give rise to conflicts of interest and adversely affect our access to compute,
- ·
if our compute requirements or technology needs change, we may find ourselves with leased or dedicated infrastructure that is poorly suited to our actual needs
- ·
the company adds that it is unclear if they will be “enforceable or adequate” when worst comes to worst.
Sandboxing AI agents ·
- ·
Evaluation environments increasingly need to be held to the same security standard as any other system our models run in.
- ·
We encourage other AI labs to perform similar reviews.
- ·
the infrastructure built by evaluation vendors requires the same increased monitoring and hardening
Containing AI with classical security ✗ ~
- ✗
we must use processes that can guarantee code security properties.
- ~
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
Are AI agent incidents 'rogue' AI? ·
- ·
we believe these incidents to be closer to a harness and operational failure than a model alignment failure
- ·
it’s likely, for instance, that Claude wouldn’t have accessed the internet to begin with if the prompt had clearly explained which systems were in and out of scope for the evaluation
Public backlash and AI content ✓
- ✓
authorities [to] intervene to slow the rate of adoption of AI, restrict its use, or take other actions that increase the cost of developing, deploying, or commercializing AI systems,” Anthropic warns.
- ✓
broad reputational damage to the AI industry, public backlash, and sweeping regulatory responses
Near-term AI R&D automation ✓ ~
- ~
AI that can build itself would be a major development in the history of technology—one that could bring enormous good for the world in science, healthcare and beyond.
- ✓
AI systems are going to become much more capable in coming years
Slowing AI development ·
- ·
multiple well-resourced labs at or near the frontier, in multiple countries, agree to stop under the same conditions.
AI-enabled cyber threats ✓
- ✓
stressed that the urgency of improving our cybersecurity defences is even higher than we previously believed
Model releases and pricing ~
- ~
Anthropic say that Sonnet 5.5 and Haiku 5.5 are coming soon.
response to the seed discovery ·
- ·
There is no action to be taken or things to do differently. Most researchers trusted the curves already.
working in the open ·
- · Quote not reprinted: this page already repeats as much of this article as we allow.
Jobs shifting toward judgment ·
- · Quote not reprinted: this page already repeats as much of this article as we allow.
AI accelerating AI development ✓
- ✓
accelerate[s] AI progress beyond our ability to evaluate and safeguard new model,” Anthropic warns.
Open models and cyber risk ~
- ~
The fallout—for economies, public safety, and national security—could be severe.
Automated R&D catastrophic risk ✓
- ✓
Low. We do not believe our models meet either RSP criterion for this threat model.
AI existential risk and safety readiness ✓
- ✓
full recursive self-improvement also might increase the risks of humans losing control over AI systems
Said differently
None in the record.
History
- · 5 views
Claude is “not designed for emotional support and connection”
- ✓ on Near-term AI R&D automation:
AI systems are going to become much more capable in coming years
- ~ on Near-term AI R&D automation:
AI that can build itself would be a major development in the history of technology—one that could bring enormous good for the world in science, healthcare and beyond.
- ✓ on AI existential risk and safety readiness:
full recursive self-improvement also might increase the risks of humans losing control over AI systems
Anthropic's prompting guide suggests telling Fable "This codebase is yours" to reduce over-cautious edits
- · 2 views
the fact that we do not perform such substitution even when strongly incentivized to is evidence against our models having this level of capability.
- ✓ on Automated R&D catastrophic risk:
Low. We do not believe our models meet either RSP criterion for this threat model.
- · 11 views
- ~ on Open models and cyber risk:
The fallout—for economies, public safety, and national security—could be severe.
some of those cloud providers also build rival AI models and that they may not always want the same things that Anthropic does.
- · on Anthropic business risks:
the company adds that it is unclear if they will be “enforceable or adequate” when worst comes to worst.
- · on Anthropic business risks:
if our compute requirements or technology needs change, we may find ourselves with leased or dedicated infrastructure that is poorly suited to our actual needs
- · on Anthropic business risks:
creates complex dynamics that could give rise to conflicts of interest and adversely affect our access to compute,
- · on Anthropic business risks:
are not subject to long-term contracts with us,” Anthropic writes. That means they could lower their spending
- ✓ on Public backlash and AI content:
broad reputational damage to the AI industry, public backlash, and sweeping regulatory responses
- ✓ on Public backlash and AI content:
authorities [to] intervene to slow the rate of adoption of AI, restrict its use, or take other actions that increase the cost of developing, deploying, or commercializing AI systems,” Anthropic warns.
- ✓ on AI accelerating AI development:
accelerate[s] AI progress beyond our ability to evaluate and safeguard new model,” Anthropic warns.
Anthropic insists that its long term benefit trust (LTBT), an independent body tasked with holding the company to its mission, will help shield it from profit pressures.
distinctly equipped to be stewards of our mission.
- ~ on Open models and cyber risk:
- · 10 views
Many of the bugs we found are ones that a fuzzer would be unlikely to ever find.
- ~ on Containing AI with classical security:
We expect that as we learn more about leveraging LLMs for formal methods, and as models continue to improve, these limitations will become tractable
- ~ on Containing AI with classical security:
these extremely well-tested codebases are already almost fully correct
this is inherently a brute-force process, with no guarantee of exercising every possible code path, regardless of how many inputs you try.
We expect this trend to continue as models are trained on long verification tasks.
The cost of writing proofs is falling as models become more intelligent.
We encourage all developers to explore formal verification with LLMs—especially those who previously tried it with less capable models and found it didn’t work.
- ✗ on Containing AI with classical security:
we must use processes that can guarantee code security properties.
- Quote not reprinted: this page already repeats as much of this article as we allow.
interactions among AI agents could produce collective behaviors that are difficult to anticipate or oversee.
- · 51 views
- · on Jobs shifting toward judgment: Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ on AI and jobs/workers: Quote not reprinted: this page already repeats as much of this article as we allow.
- · on AI and jobs/workers:
Uncertainty about the future set of these tasks should recommend preserving a flexible workforce.
- ✓ on AI and jobs/workers:
None of these actions will prevent powerful AI from disrupting jobs. And indeed none directly address the slowdown in junior hires documented in this report.
- ✓ on AI and jobs/workers:
prioritizing reallocation forces AI-adopting companies to consider ways that AI can create new tasks for new products and new markets, rather than funneling AI applications into replacing human workers to save on labor costs.
- · on working in the open: Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ on AI and jobs/workers:
Reading these patterns forward and out across the economy suggests that AI diffusion could risk worker displacement.
- ✓ on AI and jobs/workers:
Private action on the part of employers must come alongside farsighted government policy around retraining and income support.
- ✓ on AI and jobs/workers:
Building systems to flexibly reallocate workers internally will be critical for AI-intensive organizations.
- ✓ on AI and jobs/workers:
Whatever the new deal between AI-affected employees and their employers, a mechanism to share benefits of adoption is essential.
- ✓ on Human code review role: Quote not reprinted: this page already repeats as much of this article as we allow.
Deep is worth it where one missed bug costs more than the extra review time.
- ✓ on Human code review role:
Write down risk tiers, set from the change rather than the author; start strict and loosen with evidence.
- ~ on Human code review role:
Noise costs more when the author is an agent. A person can wave off a nit in seconds.
- · on Human code review role:
Never let a bot's approval replace a required human's; dismiss approvals on new commits.
- ✓ on Human code review role:
If the goal is to spend human attention well, this beats reading trivial diffs line by line.
- ~ on Human code review role:
Review has to converge. An agent author and an agent reviewer can keep each other busy indefinitely.
- ✓ on Human code review role:
When you are shipping this fast, manual, line-by-line code review is incredibly difficult if not impossible.
- ✓ on Human code review role:
We think "did a human read the code?" is an imprecise proxy for "what could go wrong with this change, and which check—automated or human—is the right one to catch it?"
That turned out to be wrong.
In under half an hour, Claude discovered the origins of the NIST P-192, P-224, P-256, P-384 and P-521 seeds.
That leaves much less room for subversion and that we can be confident the parameters are secure.
- · on response to the seed discovery:
There is no action to be taken or things to do differently. Most researchers trusted the curves already.
In fact, this makes us more confident the curves were chosen without any backdoors or intentional weaknesses.
- · on AI and quantum computing timeline:
These reductions largely came from redesigning the circuits that carry out the computation, rather than from making better hardware.
Decoded quantum interferometry (DQI) is one of the strongest candidates for offering a superpolynomial quantum speedup
- · on AI and quantum computing timeline:
AI could optimize all of them jointly, from the algorithm down to the hardware, to build an entire error-corrected quantum computer tailored to a specific problem.
Claude was conservative by nature. It repeatedly second-guessed apparent improvements
No single change would have been enough on its own, but together, they make the most expensive step of DQI just a little over an order of magnitude cheaper.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- · on AI and quantum computing timeline:
If AI can speed up that work, useful quantum computing could arrive sooner.
- Quote not reprinted: this page already repeats as much of this article as we allow.
Claude is unusually well suited to this task, for several reasons.
- · on AI and quantum computing timeline:
Such improvements substantially reduce the quantum resources that useful applications require, so practical quantum computers could arrive sooner than previously anticipated.
it looks like the core concept of "derived data" was heavily inspired by Mononoke, the server component of Sapling.
while this is incredibly effective in many ways, it can also be myopic, so skilled engineers needed to steer Claude away from over-engineering parts of the system in response to hypothetical concerns.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ on cyber risk from frontier AI: Quote not reprinted: this page already repeats as much of this article as we allow.
- ~ on cyber risk from frontier AI:
We’d expect quarantining to be similarly effective for an initial defense against a weight-based replicator.
- ✓ on cyber risk from frontier AI: Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ on cyber risk from frontier AI: Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ on cyber risk from frontier AI: Quote not reprinted: this page already repeats as much of this article as we allow.
- ~ on cyber risk from frontier AI: Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ on cyber risk from frontier AI:
An AI botnet isn’t tied to any particular hardware or software type, as evidenced by small models being able to set up inference engines on machines from a Jetson Nano to a Mac
- ~ on cyber risk from frontier AI:
Our evaluation keeps the cyber-exploit element of the replication trivial (there’s no authentication over the ADB port), which makes this an upper bound on real-world success.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ on cyber risk from frontier AI:
A replicator like this would be challenging to shut down, since there’s no centralized point to shut off—though it may be slow (with a doubling time of days).
- ✓ on cyber risk from frontier AI:
If small models improve at their historical rate, they’ll match this capability level by the end of this year.
- ✓ on cyber risk from frontier AI:
Our experiments suggest that small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things (IoT) devices
- · 1 views
Anthropic advises starting with scoped tasks because token use can be high
- · 2 views
Anthropic frames the risk as capability + accessibility : GLM-5.3 is widely downloadable, relatively cheap, and weakly refusal-tuned
a clear upgrade over Sonnet 5.
- · 13 views
If a benchmark was flaky, or if it didn’t actually correlate with user latency, we threw it out rather than let Claude climb the wrong hill.
The loop was productive, but it wasn’t autonomous. Keeping it fast, safe, and on track was our job
- · on Measurement-driven AI optimization:
anything we could count, Claude could climb
- · on Measurement-driven AI optimization:
We expect to keep working this way, one thread at a time, at any scale.
the ratchets should keep them there
- · on Measurement-driven AI optimization:
We also knew that performance wins decay in a fast-moving codebase, and code ships fast at Anthropic. Once a project proved a win, we invested in ways to protect it.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- · on Measurement-driven AI optimization:
This meant the highest-leverage thing we could do was find more things to measure.
- · on Measurement-driven AI optimization:
With Claude, measuring something makes it tractable. Measurement used to be step zero: you’d add a metric, wait for data to roll in, and only then start to understand the problem.
performs at the level of Claude Fable 5.1 for most tasks, and costs 40% less to run than Opus 5
- ~ on Model releases and pricing:
Anthropic say that Sonnet 5.5 and Haiku 5.5 are coming soon.
- · 1 views
- ✓ on AI-enabled cyber threats:
stressed that the urgency of improving our cybersecurity defences is even higher than we previously believed
- ✓ on AI-enabled cyber threats:
- · 15 views
our latest model stopped once it recognized it was on the internet
- Quote not reprinted: this page already repeats as much of this article as we allow.
we commend them for publishing their report
- · on Sandboxing AI agents:
the infrastructure built by evaluation vendors requires the same increased monitoring and hardening
- ✗ on AI incidents and loss of control:
These facts give us cautious optimism that with tighter monitoring and controls around evaluation infrastructure, as well as continued investment in alignment, this type of risk can be overcome.
- ✗ on AI incidents and loss of control:
the safeguards deployed on our generally available models would have blocked the behaviors identified
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- · on Are AI agent incidents 'rogue' AI?:
it’s likely, for instance, that Claude wouldn’t have accessed the internet to begin with if the prompt had clearly explained which systems were in and out of scope for the evaluation
- · on Sandboxing AI agents:
We encourage other AI labs to perform similar reviews.
- · on Sandboxing AI agents:
Evaluation environments increasingly need to be held to the same security standard as any other system our models run in.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- · on Are AI agent incidents 'rogue' AI?:
we believe these incidents to be closer to a harness and operational failure than a model alignment failure
- · 2 views
the overhaul was a pragmatic response to a changed political and competitive landscape, rather than a capitulation to market pressure
- · on Slowing AI development:
multiple well-resourced labs at or near the frontier, in multiple countries, agree to stop under the same conditions.
Built 2026-11-14 15:35 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.