Original post
Brendan Dolan-Gavitt#1386
Brendan Dolan-Gavitt#1386Anthropic releases initial Project Glasswing report showing its unreleased Mythos Preview models found over 10,000 vulnerabilities and blocked a $1.5 million fraud attempt while outperforming GPT-5.5 on exploit benchmarks
AI Judge changed title after evaluation, original title: "Anthropic has operated Mythos Preview and Mythos-class models internally for three months as part of Project Glasswing amid insufficient safeguards for public release"
Anthropic will expand access to more US and allied government partners first.
1.4K17.8K1.2K3.3K4M
Sentiment
Positive users praise Anthropic's Mythos models and Project Glasswing for strong benchmark results on cyber exploits, while negative users dismiss the claims as fake or criticize high costs and manipulative practices.
Pos
36.0%
Neg
64.0%
310 comments with sentiment.
Cluster Engagement
Digg Deeper
No Digg Deeper questions have been answered for this story yet.
Posts from X
Most Activity
Most Activity
VIEWS261.4K
Lisan al Gaib@scaling01
I don't understand how people are still coping about Mythos.
Here's a few benchmarks: SWE-bench Pro: Mythos -> 77.8%, GPT-5.5 -> 58.6% HLE: Mythos -> 56.8%, GPT-5.5 -> 41.4%
UK AISI cyber ranges: - "The Last Ones": Mythos -> 6/10, GPT-5.5 3/10 - "Cooling Tower": Mythos -> 3/10, GPT-5.5 0/10
ExploitBench: - Mythos -> 18 Arbitrary Code Executions - GPT-5.5 -> 0 Arbitrary Code Executions
ExploitGym: - Mythos -> 157 exploits (289.3 LLM calls) - GPT-5.5 -> 120 exploits (375.4 LLM calls)
XBOW same story. Mythos has much higher odds of finding vulnerabilities within smaller token budgets.
Ravid Shwartz Ziv@ziv_ravid
Anthropic isn't releasing Mythos. The Official reason is that it's too dangerous and could be used to exploit zero-days at scale.
Honest poll: how many of you think that if Anthropic had the compute to serve Mythos to everyone, they would still be holding it back?
Quite the coincidence that safety narratives and compute constraints have started to rhyme so perfectly, no?
38dViews 261.4KLikes 475Bookmarks 105
BOOKMARKS337LIKES1.5K
alice@aliceisplaying
oh it's bad bad
Anthropic@AnthropicAI
Last month we launched Project Glasswing, our collaborative AI cybersecurity initiative. Since then, we and our partners have found more than ten thousand high- or critical-severity vulnerabilities in essential software.
38dViews 174.3KLikes 1.5KBookmarks 337
RETWEETS538
Anthropic@AnthropicAI
Last month we launched Project Glasswing, our collaborative AI cybersecurity initiative. Since then, we and our partners have found more than ten thousand high- or critical-severity vulnerabilities in essential software.
38dViews 2.6MLikes 8.4KBookmarks 1.2K
REPLIES75
Lisan al Gaib@scaling01
Claude Mythos absolutely destroys GPT-5.5 in ExploitBench and ExploitGym
Mythos finds 18 arbitrary code execution exploits GPT-5.5 finds 0
Lisan al Gaib@scaling01
I don't understand how people are still coping about Mythos.
Here's a few benchmarks: SWE-bench Pro: Mythos -> 77.8%, GPT-5.5 -> 58.6% HLE: Mythos -> 56.8%, GPT-5.5 -> 41.4%
UK AISI cyber ranges: - "The Last Ones": Mythos -> 6/10, GPT-5.5 3/10 - "Cooling Tower": Mythos -> 3/10, GPT-5.5 0/10
ExploitBench: - Mythos -> 18 Arbitrary Code Executions - GPT-5.5 -> 0 Arbitrary Code Executions
ExploitGym: - Mythos -> 157 exploits (289.3 LLM calls) - GPT-5.5 -> 120 exploits (375.4 LLM calls)
XBOW same story. Mythos has much higher odds of finding vulnerabilities within smaller token budgets.
38dViews 124.6KLikes 1.1KBookmarks 221
Aaron Levie@levie
Here’s a key line in this mythos update. This is precisely an example of why engineers don’t go away, ever.
We’ve made it far easier to create and find security issues, which means the new bottleneck is our ability to actually review, respond to, and fix the issues.
Far from AI magically solving all of this, there still is major triage work and human judgment required to do the follow on work to actually protect systems. As a result, we’re about to enter a security engineer boom.
Jevons paradox all over again.
Anthropic@AnthropicAI
Last month we launched Project Glasswing, our collaborative AI cybersecurity initiative. Since then, we and our partners have found more than ten thousand high- or critical-severity vulnerabilities in essential software.
38dViews 122.6KLikes 601Bookmarks 242
Anthropic@AnthropicAI
Patching these vulnerabilities will make us safer. But the software industry will need to adapt to the volume of vulnerabilities that models like Claude Mythos Preview will be able to find.
We discuss this in our initial update on Project Glasswing: https://www.anthropic.com/research/glasswing-initial-update
38dViews 38.1KLikes 426Bookmarks 94
Lisan al Gaib@scaling01
The compute story is cope. The gatekeeping story is cope.
Mythos is genuinely much stronger than anything we've seen so far, and if Anthropic simply let it loose instead of starting Project Glasswing there would be millions-billions of dollars in damages.
Lisan al Gaib@scaling01
I don't understand how people are still coping about Mythos.
Here's a few benchmarks: SWE-bench Pro: Mythos -> 77.8%, GPT-5.5 -> 58.6% HLE: Mythos -> 56.8%, GPT-5.5 -> 41.4%
UK AISI cyber ranges: - "The Last Ones": Mythos -> 6/10, GPT-5.5 3/10 - "Cooling Tower": Mythos -> 3/10, GPT-5.5 0/10
ExploitBench: - Mythos -> 18 Arbitrary Code Executions - GPT-5.5 -> 0 Arbitrary Code Executions
ExploitGym: - Mythos -> 157 exploits (289.3 LLM calls) - GPT-5.5 -> 120 exploits (375.4 LLM calls)
XBOW same story. Mythos has much higher odds of finding vulnerabilities within smaller token budgets.
38dViews 29.8KLikes 397Bookmarks 64
Andrew Curran@AndrewCurran_
Update on Mythos and Project Glasswing: 'Next, we will work with critical partners—including US and allied governments—to expand Project Glasswing to additional partners. And in the near future, once we’ve developed the far stronger safeguards we need, we look forward to making Mythos-class models available through a general release.'
Anthropic@AnthropicAI
Last month we launched Project Glasswing, our collaborative AI cybersecurity initiative. Since then, we and our partners have found more than ten thousand high- or critical-severity vulnerabilities in essential software.
38dViews 30.7KLikes 397Bookmarks 53
Lisan al Gaib@scaling01
Anthropic: "once we've developed the far stronger safeguards we need, we look forward to making Mythos-class models available through a general release"
38dViews 20.3KLikes 474Bookmarks 35
Lisan al Gaib@scaling01
Anthropic is codemaxxing OpenAI is mathmaxxing
the question is, which is going to be more useful and transfers better to the other domain?
Lisan al Gaib@scaling01
I don't understand how people are still coping about Mythos.
Here's a few benchmarks: SWE-bench Pro: Mythos -> 77.8%, GPT-5.5 -> 58.6% HLE: Mythos -> 56.8%, GPT-5.5 -> 41.4%
UK AISI cyber ranges: - "The Last Ones": Mythos -> 6/10, GPT-5.5 3/10 - "Cooling Tower": Mythos -> 3/10, GPT-5.5 0/10
ExploitBench: - Mythos -> 18 Arbitrary Code Executions - GPT-5.5 -> 0 Arbitrary Code Executions
ExploitGym: - Mythos -> 157 exploits (289.3 LLM calls) - GPT-5.5 -> 120 exploits (375.4 LLM calls)
XBOW same story. Mythos has much higher odds of finding vulnerabilities within smaller token budgets.
38dViews 31.7KLikes 280Bookmarks 32
Lisan al Gaib@scaling01
doomers are going to love this sentence
btw it's been 3 months since claude mythos was deployed internally, meaning that time-horizons have almost doubled again
38dViews 11.8KLikes 307Bookmarks 31
Lisan al Gaib@scaling01
The only other lab that could have a mythos class model right now is OpenAI and if they had a model as capable as Mythos why wouldn't they do a "PR campaign"? Why would they leave the stage to Anthropic?
It's not an internal model when dozens of companies are using the model.
You are just turbo coping
sdmat@sdmat123
@scaling01 Because it's an internal model. It's not relevant, you can't use it.
The other labs have internal models too, they just don't do the PR campaign.
38dViews 27.3KLikes 263Bookmarks 28
Lisan al Gaib@scaling01
> "I can drive 4x cheaper to the next city with my lamborghini compared to your helicopter"
> "great, now get me to that island with your lamborghini"
price/performance doesn't matter if you are capability locked
Delta Vee@deltaVee42
@scaling01 Mythos is ~4x more expensive than GPT 5.5 per output token and ~5x more expensive per input token so measuring compute by tokens isn't fair. If you measure compute by API cost those "The Last Ones" and XBOW graphs imply Mythos and 5.5 would be roughly equal in capability.
38dViews 18.3KLikes 237Bookmarks 22
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
> The compute story is cope. if this were true Mythos would not have such dogshit usage limits and short context btw. They *are not* able to serve it, and they're coping when they deny it
Lisan al Gaib@scaling01
The compute story is cope. The gatekeeping story is cope.
Mythos is genuinely much stronger than anything we've seen so far, and if Anthropic simply let it loose instead of starting Project Glasswing there would be millions-billions of dollars in damages.
35dViews 11.1KLikes 151Bookmarks 14
Lisan al Gaib@scaling01
and the fact that companies are using Mythos and not GPT-5.5 for finding vulnerabilities
+ the government got involved
Lisan al Gaib@scaling01
I don't understand how people are still coping about Mythos.
Here's a few benchmarks: SWE-bench Pro: Mythos -> 77.8%, GPT-5.5 -> 58.6% HLE: Mythos -> 56.8%, GPT-5.5 -> 41.4%
UK AISI cyber ranges: - "The Last Ones": Mythos -> 6/10, GPT-5.5 3/10 - "Cooling Tower": Mythos -> 3/10, GPT-5.5 0/10
ExploitBench: - Mythos -> 18 Arbitrary Code Executions - GPT-5.5 -> 0 Arbitrary Code Executions
ExploitGym: - Mythos -> 157 exploits (289.3 LLM calls) - GPT-5.5 -> 120 exploits (375.4 LLM calls)
XBOW same story. Mythos has much higher odds of finding vulnerabilities within smaller token budgets.
38dViews 10.9KLikes 146Bookmarks 14
Ravid Shwartz Ziv@ziv_ravid
Anthropic isn't releasing Mythos. The Official reason is that it's too dangerous and could be used to exploit zero-days at scale.
Honest poll: how many of you think that if Anthropic had the compute to serve Mythos to everyone, they would still be holding it back?
Quite the coincidence that safety narratives and compute constraints have started to rhyme so perfectly, no?
Anthropic@AnthropicAI
Last month we launched Project Glasswing, our collaborative AI cybersecurity initiative. Since then, we and our partners have found more than ten thousand high- or critical-severity vulnerabilities in essential software.
38dViews 65.2KLikes 95Bookmarks 13
Gary Marcus@GaryMarcus
1. Agreed w @scaling01 that Mythos appears to be better GPT 5.5 on many metrics. 2. Mythos is definitely a major wakeup call wrt security, and will pose problems for real-world systems that aren’t well-defended. As @scaling01 says elsewhere a full release of it this point would cause a huge mess. 3. The big open question is how well Mythos does in open-ended real world problems.
Lisan al Gaib@scaling01
I don't understand how people are still coping about Mythos.
Here's a few benchmarks: SWE-bench Pro: Mythos -> 77.8%, GPT-5.5 -> 58.6% HLE: Mythos -> 56.8%, GPT-5.5 -> 41.4%
UK AISI cyber ranges: - "The Last Ones": Mythos -> 6/10, GPT-5.5 3/10 - "Cooling Tower": Mythos -> 3/10, GPT-5.5 0/10
ExploitBench: - Mythos -> 18 Arbitrary Code Executions - GPT-5.5 -> 0 Arbitrary Code Executions
ExploitGym: - Mythos -> 157 exploits (289.3 LLM calls) - GPT-5.5 -> 120 exploits (375.4 LLM calls)
XBOW same story. Mythos has much higher odds of finding vulnerabilities within smaller token budgets.
37dViews 17KLikes 39Bookmarks 23
Lisan al Gaib@scaling01
even my goat gary agrees that Mythos > GPT-5.5
37dViews 9.9KLikes 165Bookmarks 6
Dawn Song@dawnsongtweets
Excited to see ExploitGym used to evaluate Claude Mythos Preview’s cyber capabilities.
ExploitGym tests a critical question: can AI agents turn real vulnerabilities into working exploits? ExploitGym is the first comprehensive benchmark to systematically evaluate this question, including 898 real-world exploitation tasks across userspace programs, V8, and the Linux kernel, with realistic mitigation toggles.
Exploitation remains challenging, but frontier models are making significant progress, underscoring the need for rigorous, reproducible, dual-use cyber evaluations.
Anthropic@AnthropicAI
Last month we launched Project Glasswing, our collaborative AI cybersecurity initiative. Since then, we and our partners have found more than ten thousand high- or critical-severity vulnerabilities in essential software.
34dViews 7.8KLikes 50Bookmarks 16
Taelin@VictorTaelin
@scaling01 reminder that this is all *without thinking*
AFAIK Mythos is a phase transition just like what we observed from GPT-2 to GPT-3. it is a whole new beast
Lisan al Gaib@scaling01
I don't understand how people are still coping about Mythos.
Here's a few benchmarks: SWE-bench Pro: Mythos -> 77.8%, GPT-5.5 -> 58.6% HLE: Mythos -> 56.8%, GPT-5.5 -> 41.4%
UK AISI cyber ranges: - "The Last Ones": Mythos -> 6/10, GPT-5.5 3/10 - "Cooling Tower": Mythos -> 3/10, GPT-5.5 0/10
ExploitBench: - Mythos -> 18 Arbitrary Code Executions - GPT-5.5 -> 0 Arbitrary Code Executions
ExploitGym: - Mythos -> 157 exploits (289.3 LLM calls) - GPT-5.5 -> 120 exploits (375.4 LLM calls)
XBOW same story. Mythos has much higher odds of finding vulnerabilities within smaller token budgets.
38dViews 8.3KLikes 101Bookmarks 8
Load more posts