SafetyRisk, alignment & guardrails
Armin Ronacher says decades of bad software are catching up as AI exposes critical PHP flaws and floods open-source maintainers
Flask creator Armin Ronacher argued that three decades of poorly built software are now coming due. His remark landed the same week AI-assisted research exposed two critical, actively exploited PHP flaws and AI-generated bug reports overwhelmed curl, OpenSSL and Google's bounty program.
From the Simon Willison article of October 22, 2026 and cited sources: Emergency PHP releases fixed two flaws, CVE-2026-109618 and CVE-2026-109619, both rated CVSS 9.2. [3] [7] [8]
SOURCES[3] Urgent security fix issued for code vital to estimated 70%… abc7news.com[7] Techmeme, Oct 21, 2026 techmeme.com[8] OpenAI's Codex Security helps uncover two critical PHP… cryptobriefing.com
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Emergency PHP releases fixed two flaws, CVE-2026-109618 and CVE-2026-109619, both rated CVSS 9.2. [3] [7] [8]
- Original source
- Simon Willison article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 22, 2026 article published
- Covered
- Oct 25 edition
- Evidence
- 13 sources across 13 sites
- 1A quote from Armin Ronacher – Simon Willisonsimonwillison.net
- 2Armin Ronacher on Xx.com
- 3Urgent security fix issued for code vital to estimated 70%…abc7news.com
- 4AI slop submissions force Google to freeze its open-source…helpnetsecurity.com
- 5curl security advisories – oss-securityopenwall.com
- 6Swamped curl maintainers… – The Registertheregister.com
Reach: not measured yet
Emergency PHP releases fixed two flaws, CVE-2026-109618 and CVE-2026-109619, both rated CVSS 9.2. They are remote code execution bugs that need no authentication. They stem from how PHP handles out-of-memory conditions and from global state that persists across requests, and they affect versions back to PHP 5.4.0. The security firm Calif reportedly found one using an OpenAI model, and Wiz found both independently. An exploitation attempt was detected before the fixes shipped, and CISA and the FBI urged immediate patching. [3] [7] [8]
AI-generated bug reports are changing how open-source projects take in vulnerability reports. Google froze new submissions to its open-source bounty program, citing a surge of mostly invalid automated reports, and expects a revised program in Q1 2027. curl fixed 24 CVEs in one release, only one of them rated HIGH, and will handle low- and medium-severity reports in public. OpenSSL's security maintainers report similar strain. [4] [5] [6] [12] [13]
PHP design flaws decades old were exploited around disclosure, and maintainers are changing how they take in reports because of the volume.
Read the full assessment
Implication: teams should expect more CVEs in old dependencies, plan patch capacity to match, and review agent-written code for maintainability as well as task completion.
Executive brief
Armin Ronacher, the creator of Flask, posted on X that "we built so much bad software over the last 30 years, and now the universe is punishing us for our sins." Simon Willison quoted it on October 22 under his security and AI-security-research tags. Neither page says which event Ronacher was reacting to. The same week, though, AI-assisted research turned up two critical remote-code-execution flaws in PHP, which runs on roughly 70% of websites, and exploitation was reported within hours of the patches. Meanwhile AI-generated bug reports are overwhelming curl, OpenSSL and Google's open-source bounty program.
What changed and event timeline
Google freezes its open-source bounty
Google stopped taking new OSS VRP submissions, citing a rise in automated reports that are mostly invalid. It promised a revised program in Q1 2027 ().
curl ships 24 CVEs in one release
curl 8.23.0 fixed 24 vulnerabilities. One was rated HIGH, an HTTP/2 upload use-after-free; the rest were LOW ().
Exploitation attempt before patches
A first PHP exploitation attempt was detected on Amazon's systems, ahead of the public fixes ().
curl takes low- and medium-severity reports public
Starting October 26, curl will handle low and medium reports in public on GitHub. Maintainer Daniel Stenberg described the volume as an "avalanche" ().
PHP emergency releases
PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35 fixed CVE-2026-109618 and CVE-2026-109619, both rated CVSS 9.2. The release reportedly came a day early because details had leaked (). ABC7 gives the patch date as October 22.
Ronacher's remark
Willison quoted Ronacher's post. Neither gives an explicit trigger ().
Capabilities and access
- PHP flaws: the security firm Calif reportedly found one of them with OpenAI's GPT-5.6 Sol under OpenAI's "Patch the Planet" initiative, and Wiz found both independently (Techmeme).
- Anthropic: Claude Fable 5 (formerly Mythos) reportedly surfaced 6,202 high or critical findings out of 23,019 total, across more than 1,000 projects (Atlantic Council).
- Ronacher's own test: in September he ran GPT-6 Astra as an autonomous agent for 35 hours (LavX).
Read the full section
- PHP flaws: the security firm Calif reportedly found one of them with OpenAI's GPT-5.6 Sol under OpenAI's "Patch the Planet" initiative, and Wiz found both independently (Techmeme). Another outlet credits OpenAI Codex Security alongside Calif and Wiz (Crypto Briefing).
- Anthropic: Claude Fable 5 (formerly Mythos) reportedly surfaced 6,202 high or critical findings out of 23,019 total, across more than 1,000 projects (Atlantic Council).
- Ronacher's own test: in September he ran GPT-6 Astra as an autonomous agent for 35 hours (LavX).
Technical analysis for researchers and developers
- Root causes: the PHP bugs come from how PHP handles out-of-memory conditions and from global state that persists across requests.
- Validation: remote code execution was reportedly confirmed on WordPress and Drupal in August 2026 (Crypto Briefing).
- Generated code quality: Ronacher's 35-hour run cost $1,200 and produced 75,000 lines of code he judged "objectively bad."
Read the full section
- Root causes: the PHP bugs come from how PHP handles out-of-memory conditions and from global state that persists across requests. Affected versions go back to PHP 5.4.0, and exploitation needs no authentication (Crypto Briefing). These are old design flaws, not new code, which fits the "30 years" framing.
- Validation: remote code execution was reportedly confirmed on WordPress and Drupal in August 2026 (Crypto Briefing). No public methodology for the model-driven discovery was found in the reviewed sources.
- Generated code quality: Ronacher's 35-hour run cost $1,200 and produced 75,000 lines of code he judged "objectively bad." He argues that models tuned for task completion and token efficiency write code people cannot read (LavX).
Claims and evidence
- OpenAI calls the PHP flaws one of the most impactful security issues in recent memory. This is the vendor's own characterization (Techmeme).
- Amazon reported broad in-the-wild exploitation, and CISA and the FBI urged immediate patching (ABC7).
- Report noise. More than 90% of recent curl submissions were assessed below HIGH severity (The Register).
Read the full section
- OpenAI calls the PHP flaws one of the most impactful security issues in recent memory. This is the vendor's own characterization (Techmeme). Wiz's independent discovery is the stronger support for the flaws being real.
- Active exploitation. Amazon reported broad in-the-wild exploitation, and CISA and the FBI urged immediate patching (ABC7).
- Report noise. More than 90% of recent curl submissions were assessed below HIGH severity (The Register). One blog claims report volume is up 76% while only 25% of reports are valid, but its methodology is unclear (danilchenko.dev).
- No independent corroboration was found for the Claude Fable 5 counts.
Context and prior work
- curl ended its HackerOne bounty earlier in 2026 because of AI-generated submissions. Intel also dropped bounty payments (Malwarebytes).
- In June 2026 the Atlantic Council urged frontier labs to fund maintainers' triage work and to show restraint in vulnerability discovery.
- Axios reports that AI is overwhelming OpenSSL's security maintainers (Axios).
Read the full section
- curl ended its HackerOne bounty earlier in 2026 because of AI-generated submissions. Intel also dropped bounty payments (Malwarebytes).
- In June 2026 the Atlantic Council urged frontier labs to fund maintainers' triage work and to show restraint in vulnerability discovery. It also noted that supply-chain attacks such as Shai-Hulud targeted maintainers' infrastructure, not code bugs (Atlantic Council).
- Axios reports that AI is overwhelming OpenSSL's security maintainers (Axios).
Limitations, safety and contested findings
- Contested patch date: October 21 per Techmeme, October 22 per ABC7.
- Disclosure process: the PHP Foundation called the delay "a structural failure," and exploitation attempts came before the fixes (ABC7).
- Trade-off at curl: taking reports public spreads triage work across more people, but gives attackers more time before fixes ship (The Register).
Read the full section
- Contested patch date: October 21 per Techmeme, October 22 per ABC7.
- Disclosure process: the PHP Foundation called the delay "a structural failure," and exploitation attempts came before the fixes (ABC7).
- Trade-off at curl: taking reports public spreads triage work across more people, but gives attackers more time before fixes ship (The Register).
- Inferred link: connecting Ronacher's remark to these events is an inference from timing and Willison's tags.
Business and practitioner implications
- Cloudflare turned on WAF rules about nine hours before disclosure, but stopgap mitigations reportedly do not fully block exploitation (Techmeme, ABC7).
- Plan for more disclosures: expect more CVEs in old dependencies and plan patch capacity to match.
- Report quality: if you run a bounty program, expect AI pre-filtering. Google is moving toward escalating only validated findings (Malwarebytes).
Read the full section
- Patch now: update PHP immediately. Cloudflare turned on WAF rules about nine hours before disclosure, but stopgap mitigations reportedly do not fully block exploitation (Techmeme, ABC7).
- Plan for more disclosures: expect more CVEs in old dependencies and plan patch capacity to match.
- Report quality: if you run a bounty program, expect AI pre-filtering. Google is moving toward escalating only validated findings (Malwarebytes).
- Agent-written code: review it for readability and maintainability, not just whether the task passes (LavX).