Study says AI-written patches are often incomplete fixes, instead becoming attack roadmaps
Start here
Multiple studies and security conference reports show that AI-written patches are often incomplete fixes, instead marking vulnerability locations and defensive approaches for attackers. Of the 6,080 AI patch candidates tested by 1Password Off-by-1 Labs, about half left at least one attack path; in the P2Patch benchmark, attack agents successfully bypassed patches in 41 of 101 cases. Anthropic researchers, using only public information from patch day, produced a working exploit within one hour of the Firefox fix landing; a Fudan University team's LLM-assisted binary diffing framework confirmed 26 bypassable patches out of 62 closed-source applications' patches, obtaining 19 CVEs. Epoch AI researchers used GLM 5.3 to simulate an ordinary attacker and, spending about $1,000 in API credits, broke every Firefox version between September 17 and October 26, prompting Mozilla to decide to stop publishing regression tests alongside security fixes.
Generated by AI from reports · Updated 3 days ago
Attention and developments
Attention, hourly
Now 42, -1% in 24 hours·Peak 48 (Nov 17 01:00)
Progress timeline
- X@leisu559327· Discussion@don_tech1 Patch turns into attack blueprint(opens the original post in a new tab)
- CSO OnlineSelectedAI-written patches are becoming hackers' attack maps
Multiple studies and security conference reports show that AI-written patches are often incomplete, and instead mark out vulnerability locations and defensive approaches for attackers. Of the 6,080 AI patch candidates tested by 1Password Off-by-1 Labs, about half left at least one attack path; in the P2Patch benchmark, attack agents successfully bypassed patches in 41 of 101 cases. Anthropic researchers, using only public information from patch day, produced a working exploit within one hour of the Firefox fix landing; a Fudan University team's LLM-assisted binary comparison framework also confirmed that 26 of the patches for 62 closed-source applications could be bypassed, earning 19 CVEs. Epoch AI researchers used GLM 5.3 to simulate an ordinary attacker and, spending about $1,000 in API credits, broke every Firefox version between September 17 and October 26, leading Mozilla to decide it will no longer publish regression tests alongside security fixes.
Related discussion
- WebAI Era (Xinzhiyuan)Patches became a "vulnerability instruction manual"! Google reveals hackers had AI write attack code based on patches(opens the original post in a new tab)
- X@Fly_Sistah@JasJWright Yes. Absolutely update. Hackers are reportedly using AI to scan software updates, figure out which security…(opens the original post in a new tab)
- X@xsm0820@Fang_yuanyi The patch was reverse-engineered into attack code right after it was released—this speed is indeed frightening. Defense simply can't keep up; in the future it won't just be about who discovers vulnerabilities first, but who applies AI to defense first~(opens the original post in a new tab)
- X@Arrowuppz@wellwue The matter of patches becoming attack roadmaps is quite realistic; in the future, defenders will most likely also have to rely on AI to win the time gap(opens the original post in a new tab)
- X@leisu559327@don_tech1 Patch turns into attack blueprint(opens the original post in a new tab)
- X@david1989_zhu@DrKarte Patches becoming "attack blueprints" is not new; what is more worth watching with AI is that it shortens the time from patch analysis to exploit. The most dangerous window for defenders is when the patch is already public but has not yet been applied in production. Internet-exposed services can no longer be fixed slowly on a regular schedule.(opens the original post in a new tab)
3 likes·Nov 17 09:13Collect
- X@david1989_zhu@NewsFromGoogle What deserves the most attention here is not "hackers used LLMs" but the time gap between patch publication and completion of remediation in the production environment. Defense tools can speed up risk identification, but if change approval and canary releases still have to queue, the window for attackers to act first remains.(opens the original post in a new tab)
1 likes·Nov 17 08:13Collect
- Older than 72 hours · Not included in the 72-hour figures
- X@conradlotzAI-written patches leave residual attack paths open Researchers found AI-generated patches often fail to fully close t…(opens the original post in a new tab)
- X@limio67637193Previously, a 30-day "patch adoption period" was set aside because traditional hackers relied on manual reverse engineering of patches to find exploit chains, which often took weeks; now a large model can look at open-source commit records or binary diffs and automatically generate a usable exploit in minutes. Whether a vulnerability is public or not cannot stop an attacker with AI compute at all; continuing to keep it secret instead blinds the defender unilaterally…(opens the original post in a new tab)
- WebTechmemeGoogle's TIG reports the first observed instance of attackers using an autonomous AI framework with no human in the loop to turn security patches into exploits (Duncan Riley/SiliconANGLE)(opens the original post in a new tab)
Attention trend
New sources per day
- Nov 16: 3 new sources (2 outlets, 1 accounts)
- Nov 17: 4 new sources (0 outlets, 4 accounts)
- Nov 18: 2 new sources (0 outlets, 2 accounts)
- Nov 19: 1 new sources (1 outlets, 0 accounts)
- Nov 20: 0 new sources (0 outlets, 0 accounts)