--- title: Frontier AI now outperforms humans at binary reverse... description: Frontier AI models are increasingly capable of reverse engineering binaries without source code, matching or exceeding human performance on tasks like... canonical: https://daily.dev/posts/frontier-ai-now-outperforms-humans-at-binary-reverse-engineering-wtvbgnyct twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Frontier AI now outperforms humans at binary reverse engineering | daily.dev og:description: Frontier AI models are increasingly capable of reverse engineering binaries without source code, matching or exceeding human performance on tasks like... og:url: https://daily.dev/posts/frontier-ai-now-outperforms-humans-at-binary-reverse-engineering-wtvbgnyct og:image: https://api.daily.dev/og/posts/wTvBGnYCT.png og:image:alt: Frontier AI now outperforms humans at binary reverse engineering og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Frontier AI now outperforms humans at binary reverse engineering **[Collections](https://daily.dev/sources/collections)** · 4 min read · 1 upvotes · 0 comments ## Summary Frontier AI models are increasingly capable of reverse engineering binaries without source code, matching or exceeding human performance on tasks like vulnerability discovery and exploitation. ValsAI's SRE benchmark, introduced less than a month prior, was designed to measure this exact capability and is already being saturated by current models, raising questions about the trajectory of AI-assisted security research and offensive tooling. ## Content ## What's happening Microsoft's September security update patched roughly 972 vulnerabilities, 112 of them rated critical. That followed 570 in July and 620 in August. Google and other vendors are posting similar numbers. The reason isn't that software suddenly got worse - it's that AI is finding bugs faster than humans ever could. Researchers at Arizona State University ran 128 parallel Codex agents against the Linux kernel. After refining their workflows and how they extracted vulnerability properties, the agents surfaced over 1,000 vulnerabilities triggerable by unprivileged users. That's more than any manual reporting process could realistically handle. Meanwhile, an open-weight model - GLM 5.3-flash - can run on consumer hardware costing a few thousand dollars. Once its safety filters are removed (a process called "abliteration"), it scores extremely high on exploitation benchmarks like CyberGym and ExploitBench. This isn't a capability locked behind frontier labs anymore. It's accessible to anyone with a decent GPU. ## The patch window problem Here's the part that should worry people: AI agents can now find working exploits from a vague rumor of a vulnerability, before a patch is even public. OpenAI, Anthropic, AWS, Google, Microsoft, and over 100 other organizations recently signed an open letter warning that AI-enabled attackers are shrinking the window between disclosure and exploitation to near zero. That's fundamentally incompatible with how coordinated disclosure currently works. The whole model assumes defenders have time to prepare a patch, test it, and distribute it before attackers reverse-engineer the problem. AI can reverse-engineer an exploit from a patch almost immediately after release. The embargo period that security teams rely on may already be functionally dead. ## What this means in practice The bottleneck isn't finding vulnerabilities anymore - AI is already good at that. The bottleneck is deploying fixes across legacy systems and critical infrastructure. Hospitals, banks, power utilities, and telecoms are running software that takes months or years to patch, if it gets patched at all. The argument that we have roughly a year before even more capable open models close the gap with frontier labs feels plausible. What happens after that is harder to predict, but the current trajectory isn't comfortable. Vulnerability counts will probably keep rising before they eventually fall - the optimistic read is that AI will eventually exhaust the findable bugs. But we have to get through the period where attackers are using the same tools as defenders, and defenders are losing on deployment speed. ## What actually helps Banning models or restricting GPUs won't work - the capability is already distributed. What might actually move the needle: - Governments mandating faster patch deployment and more frequent penetration testing, especially for critical infrastructure - Companies using frontier models for defensive triage, not just waiting to be attacked - Sandboxing AI agents with scoped credentials so they can't do unlimited damage if compromised - Serious investment in memory-safe languages, fuzzing, and formal verification - the boring stuff that reduces the attack surface before bugs are found - Dramatically shortening the time between patch release and deployment ## Is cybersecurity still worth learning? Enrollment in security programs is dropping, partly because students worry AI will eliminate the field. That fear is understandable but probably wrong. Agentic AI changed security research the same way fuzzing changed it before, and static analysis before that. Each wave automated the rote work and raised the floor for what humans need to contribute. The researchers who designed those 128-agent Linux kernel pipelines weren't replaced by the agents - they built them. Human expertise still matters for designing workflows, interpreting results, and doing the work that doesn't fit neatly into a benchmark. For people who are genuinely interested in security, 2026 is a strange but not a bad time to be in the field. The problems are harder and more urgent than they've ever been. That's not nothing. ## Questions this post answers ### What is the SRE benchmark for AI reverse engineering? The SRE benchmark, released by ValsAI, tests whether an AI model can reverse engineer software using only compiled binaries, without access to source code. It measures capabilities like vulnerability discovery and exploitation. Within about a month of its release, frontier AI models were already saturating the benchmark, matching or exceeding human-level performance. _Security teams tracking AI's growing reverse-engineering ability can follow benchmark updates on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 3 discussions and 760 comments across lobsters, hackernews (as of 2026-09-20). **TL;DR:** The community largely treats the post's cybersecurity urgency as a jumping-off point for broader debates about AI risk, bioweapons, and societal collapse, with significant skepticism about the severity of near-term LLM-enabled threats but genuine concern about the attacker/defender asymmetry that cheap open models create. **Sentiment:** 15% positive · 30% mixed · 55% skeptical **The case for** - LLMs genuinely worsen the attacker/defender asymmetry because attackers only need to succeed once while defenders must be right every time, and AI dramatically lowers the skill floor for attackers. - Cheap, locally-run open models could enable persistent, automated vulnerability scanning and exploitation that was previously out of reach for most individuals. - The collation and dynamic tailoring that LLMs provide meaningfully amplifies novice capability in broad domains like software exploitation, even if narrow domains (pipe bombs) see little uplift. **The pushback** - Many commenters argue the threat is overstated, pointing out that dangerous information has long been freely available without causing widespread harm. - The defender side faces bureaucratic friction, approval chains, and legal constraints that prevent AI from being deployed as freely on defense as on offense, but this asymmetry predates LLMs. - Several commenters note that the hardware cost to run a truly capable open model is still in the $10k–$20k range, limiting the immediate accessibility claim. - The discussion drifts heavily into speculative territory (bioweapons, superintelligence, Fermi paradox) that the post itself doesn't address, suggesting the concrete security argument isn't fully landing. **By community** - lobsters (mixed): No comments were posted, so no community signal is available beyond the 25 upvotes. - hackernews (mixed): The high-point thread is lively but spends most of its energy on tangential debates about AI existential risk, bioweapons, and civil liberties rather than engaging directly with the post's patch-deployment argument, with roughly equal camps of concern and skepticism. **Hottest debate:** Whether cheap open-weight LLMs meaningfully lower the barrier to catastrophic attacks or whether determined bad actors already had access to the necessary knowledge and resources without AI assistance. **Open questions** - How quickly will hardware costs drop to the point where a truly dangerous open model is accessible on consumer-grade devices? - If defenders face approval chains and legal constraints that attackers don't, what structural changes could actually close that gap? - Does the attacker/defender asymmetry in cybersecurity fundamentally worsen with AI, or does AI strengthen both sides proportionally? **Highlights** > Any sufficiently determined individual can buy mac mini, put it under their bed, configure outside proxy via some random internet address and prompt "iterate on websites in the CT logs, one by one, try to find vulnerabilities, if you did - encrypt their data and blackmail them for this bitcoin address". And it'll work, day and night. Abliterated GLM 5.3 is much smarter than average software developer, they know a lot about information security, they can use any available exploits, they can find novel vulnerabilities and they won't say "no". This is dangerous for an average IT system which never encountered nothing worse than some wordpress GET requests. It's not the end of the world. But future will be rough. > — [vbezhenar on hackernews · 4 comments](https://news.ycombinator.com/item?id=49608279) > One argument for LLMs is that although all information on topics X, Y and Z was already available somewhere, LLMs make that information more exploitable through collation, filtering and dynamic tailoring. For a relatively narrow subject area (e.g. construction of pipe bombs) the collation is minimal, and so the filtering and tailoring probably isn't that important; a novice doesn't learn a lot more from the LLM than they would have done from a few Google searches. For a broad subject (practical creation and exploitation of software vulnerabilities), the collation is very significant and the filtering means that LLMs can empower a novice to act at a similar level as an expert. > — [gnfargbl on hackernews](https://news.ycombinator.com/item?id=49607370) > No because on the defense side you need multiple layers of approvals to change anything. If not you have an LLM making production changes that can make the posture worse, or take down services, which is also bad. Once a vulnerability is discovered however if it's in your own software a patch has to be written (without reducing functionality in most cases), tested, and deployed. At every step there will be others arguing about whether this line could do better, my service requires this thing that isn't included. So at every step the patch can be delayed. And if it is someone else's software you will be lucky if it's open source and you can write a patch yourself. If it's closed source or a vendor you have to completely rely on them and use whatever your account rep can pull. Attackers have a massive advantage with AI, partially because the defensive side doesn't want to make their side worse by giving a ln LLM admin access to all their data > — [kswagage on hackernews](https://news.ycombinator.com/item?id=49608563) > Yes, and... Not only the quantity of people who have the minimum aptitude required, specialized expertise requires both knowledge and experience doing these tasks. Using an LLM requires neither. Leaning on an LLM to do much or all of this means it can happen in seconds/minutes/hours, the LLM can do it several times during that psychotic break (as opposed to a fraction of a hack in a single episode). The barrier to entry pre-LLM was both high and the population who could pull it off (before the Chinese/Russians turned this into commerce) was low. Hacking is an VERY asymmetric activity (the attacker only needs to "be right" once, whereas the defender has to be right every time for every asset they defend). It takes geometrically / exponentially more work to defend (while keeping high availability) than it does to defend. The more widespread tools to find vulns / generate exploits are, the faster the posture of the defense side falls from "maybe we can stop most hacks" to "we know we will fail to prevent most breaches, so we need to prioritize securing only the most valuable resources". That's a BAD place for the average company to be in. > — [thephyber on hackernews](https://news.ycombinator.com/item?id=49608924) > Yes but at least with cybersecurity it does not only benefit attackers. Defenders also benefit greatly from AI. There is the worry of the old saying "they (the attackers) only have to succeed once to win, we (the defenders) only have to fail once to lose.". In that sense there is a big imbalance, but the emergence of AI does not really affect that because it strengthens both sides. With physical security like things like pipe bombs that's a lot more imbalanced. However what can we do? The only effective measures include monitoring everyone which is not a solution because it will make the world not worth living in. > — [wolvoleo on hackernews](https://news.ycombinator.com/item?id=49607319) **Source threads** - [lobsters](https://lobste.rs/s/re9wk8/we_have_year_fix_security_everywhere) · 25 points · 24 comments - [hackernews](https://news.ycombinator.com/item?id=49605691) · 312 points · 736 comments - [hackernews](https://news.ycombinator.com/item?id=49572569) · 3 points · 0 comments ## Similar posts on daily.dev - [Vulnerability Research Is Cooked — Quarrelsome](https://daily.dev/posts/vulnerability-research-is-cooked-quarrelsome-qbluzvaij) · Hacker News · 1 upvotes · 0 comments - [AI gets good at finding bugs, not as good at fixing them](https://daily.dev/posts/ai-gets-good-at-finding-bugs-not-as-good-at-fixing-them-sptd5ofrp) · The Register · 0 upvotes · 0 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#vulnerability](https://daily.dev/tags/vulnerability), [#ai-security](https://daily.dev/tags/ai-security), [#reverse-engineering](https://daily.dev/tags/reverse-engineering) [View this post on daily.dev](https://daily.dev/posts/frontier-ai-now-outperforms-humans-at-binary-reverse-engineering-wtvbgnyct) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Frontier AI now outperforms humans at binary reverse engineering","url":"https://daily.dev/posts/frontier-ai-now-outperforms-humans-at-binary-reverse-engineering-wtvbgnyct","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/frontier-ai-now-outperforms-humans-at-binary-reverse-engineering-wtvbgnyct"},"datePublished":"2026-09-07T17:28:33.740Z","dateModified":"2026-09-20T04:29:36.680Z","description":"Frontier AI models are increasingly capable of reverse engineering binaries without source code, matching or exceeding human performance on tasks like...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/frontier-ai-now-outperforms-humans-at-binary-reverse-engineering-wtvbgnyct","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,vulnerability,ai-security,reverse-engineering","timeRequired":"PT4M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Frontier AI now outperforms humans at binary reverse engineering"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/frontier-ai-now-outperforms-humans-at-binary-reverse-engineering-wtvbgnyct#faq","mainEntity":[{"@type":"Question","name":"What is the SRE benchmark for AI reverse engineering?","acceptedAnswer":{"@type":"Answer","text":"The SRE benchmark, released by ValsAI, tests whether an AI model can reverse engineer software using only compiled binaries, without access to source code. It measures capabilities like vulnerability discovery and exploitation. Within about a month of its release, frontier AI models were already saturating the benchmark, matching or exceeding human-level performance. Security teams tracking AI's growing reverse-engineering ability can follow benchmark updates on daily.dev."}}]} ```