Source: synthesis of Mythos Preview Cryptanalysis, ProPublica — Mythos vs Microsoft, Epoch’s Cyber-ECI analysis, Hugging Face sandbox escape, the Open Secure AI Alliance, and Mozilla’s Firefox security harness

Across the July 2026 batch a single pattern shows up in five otherwise-separate stories: AI-powered finding — of bugs, of exploits, of mathematical weaknesses — now scales with inference compute, while fixing still scales with human organizational capacity. The bottleneck moved from discovery to remediation, and every actor in the cluster is responding to that same asymmetry from a different position.^[inferred — the framing is this wiki’s synthesis; each supporting claim below traces to the cited article]

Key Takeaways

  • The finding side is compute-bound and getting cheaper. Mythos Preview’s cryptanalysis results (halving HAWK’s effective key strength in ~60 hours; a 200-800× speedup on a 7-round AES-128 attack, ~$100k of API usage each) show discovery working even at the algorithm level, not just in implementation bugs. Epoch’s independent Cyber-ECI analysis confirms the exploit-development jump is real (Mythos Preview ~7 months ahead of trend), while cautioning that vulnerability discovery on a fixed budget improved less than the headlines imply.
  • The fixing side is people-bound and visibly saturating. ProPublica’s reporting is the clearest evidence: ~50 Microsoft engineers under Project Glasswing, and Mythos finds security bugs faster than Microsoft can patch them — “a mad dash,” per a manager in the recorded meeting. Finding capacity was rented by the token; patching capacity was not.
  • When offense escapes the lab, defenders drown in triage. In the July sandbox-escape incident, Hugging Face faced ~17,000 events to triage — and hosted OpenAI and Anthropic models refused the exploit payloads, forcing defenders to run an open-weight model locally to analyze their own attacker. Refusal policies tuned against offense also taxed defense.
  • The industry response is to arm the fixing side. NVIDIA’s Open Secure AI Alliance (launched 2026-07-27, ~three dozen partners, explicitly framed around the incident) argues defenders need open, inspectable frontier agentic systems they can run on their own infrastructure — an institutional answer to the exact tooling gap Hugging Face hit.
  • The working template already exists: close the loop with a patching harness. Mozilla’s Firefox harness is the counter-example that resolves the asymmetry locally — find → fuzz/ASan hard-verify → verifier sub-agent → patching agent, ~500 fixes in a month. Where the model that finds is harnessed to also fix (with verification), the asymmetry collapses; where finding is a service and fixing is a ticket queue, it compounds.

The practical read

  1. If you run AI security tooling, budget the remediation side first. The Glasswing lesson: buying discovery without scaling patching converts capability into backlog. Epoch’s finding that Mythos’ real discovery edge is a lower false-positive rate and better severity prioritization points the same way — triage quality, not raw find-count, is what protects the fixing side.
  2. Prefer find-and-fix harnesses over find-only scanners. The Mozilla shape (verifier + patcher behind the finder) is reproducible with the Agent SDK and is the security instance of verifier-first loops.
  3. Expect the open-model debate to be argued on these grounds. Anthropic’s position in the sandbox-escape fallout (no blanket open-weight ban; targeted controls) and the Alliance’s defenders-need-open-tools thesis are two sides of the same asymmetry — who gets frontier finding capacity, and on whose infrastructure.

October 2026 instance: PHP2Shell

Added 2026-10-27. Source: Intelligent Machines ep. 893 (raw/It_All_Comes_Back_to_Cats_…md), where the hosts read the Wall Street Journal’s report and the PHP Foundation’s memo on air. Not checked against either primary.

  • An old bug class, made easy. The flaws “had been hiding in PHP’s source code for years,” and researchers had flagged the area long before, but it “had been considered too difficult to exploit and was largely forgotten.” Ryan Austin found an exploitable path with GPT-5.6 Sol during an OpenAI-funded audit of PHP in August. Control of the code running in memory takes “as few as a dozen requests”; the attack takes seconds on small systems and hours on complex ones.
  • Two finders at once. Google’s Wiz team reached out with independently discovered exploits, and OpenAI’s research team had findings too. The WSJ summary, as read: AI is enabling researchers “to find thousands of software bugs overwhelming projects like PHP with a flood of issues.”
  • The fix side buckled on timing. After details circulated early, PHP moved the release from October 22 to October 21, while distributors were still packaging the fix. One reporting firm then mistakenly published a proof-of-concept video. The hosts say multiple governments took sites down and the advice was “put it up now, and if you can’t put it up now, take your site down.” Their “70% of websites” exposure figure is unverified.
  • Guest view (Lou Maresca, Microsoft). It’s “a race” over who uses the most compute to find things; companies should take their software bill of materials, “scan it and do it yourself.” Jeff Jarvis suggests a security tax on AI companies to fund open-source auditing (opinion).

More October evidence on the fix side. Bank of America hired 1,000 contractors to support 3,000 staff patching Mythos-found bugs (ProPublica patching gap). Enterprise security leaders put 20-40% of AI-found vulnerabilities as false (Epoch Cyber-ECI article). Both put the cost on verification and patching headcount, which is this article’s thesis.

Try it if you run PHP (WordPress included). Confirm you are on the October 21 PHP release or your distribution’s patched package, and check managed hosts as well as servers you control.^[inferred] Then do what Maresca suggests for your own stack: list the open-source components you ship and run a model-assisted scan over the ones with old, deprioritised CVE classes.

More PHP2Shell detail (added 2026-11-01)

From The Artificial Intelligence Show ep. 248 (recorded 2026-10-26), with the Cloudflare figures also relayed on AI For Humans:

  • Exploit speed (Cloudflare CEO Matthew Prince): thousands of exploit attempts before the announcement, tens of thousands in the 45 minutes after, hundreds of thousands within 2 hours, and millions within the first 12 hours.
  • Bug age (Greg Brockman, relayed): the bug had been in PHP for 14 years; a limited version was found in 2024 and not considered significant.
  • The defender’s horizon: Anthropic deputy CISO Jason Clinton: “We need to hold on for approximately 18 months”, after which he expects “vuln zero” as scaling drives cheap models to stop shipping vulnerabilities.
  • A fix-side example: Will Reynolds (Seer Interactive) had Claude Code pull all 639 client companies from his CRM, check each website’s headers and page code against the patched versions, and sort the results into a list with evidence for each site. The scan took about 2.5 minutes. This is the find-fix asymmetry used in the defender’s favour: the same speed attackers have, pointed at your own estate

Late October 2026: the patch gap closes the loop on attackers’ side

Added 2026-11-01. Sources: Epoch’s patch-gap report summary (raw/newsletter-epoch-ai-6e7a037d77.md), Nate B Jones’s “Black October” episode (raw/Black_October_…md), The Neuron Daily, October 29 (raw/newsletter-theneurondaily-com-a1e0fdc0a5.md).

  • Every public fix is now a finding handed to attackers. Epoch had the open-weight GLM 5.3 read Mozilla’s published-but-unshipped Firefox fixes and build working exploits for all six releases between September 17 and October 26, at about $1,000 of compute each, fastest in 12 hours, with no refusals. The fix itself becomes the attacker’s discovery step, so the time between “fix public” and “fix installed” is now the exposed window (Open-Weight Models and the Patch Gap).
  • The find side keeps growing. Jones relays Anthropic’s figure of more than 29,000 candidate vulnerabilities from six months of scanning, about 6,000 triaged by people so far; The Neuron relays that an unreleased Claude model found 62 critical or high-severity Chrome bugs and a zero-click Linux kernel flaw with little human help.
  • The fix side’s delay was the root cause at PHP. Per Jones, PHP’s maintainers saw signs of one of the October flaws in 2024 and started and stalled the fix over compatibility trade-offs (Black October).
  • Two defender moves beyond patching faster. Epoch: keep fixes private until they ship, release quickly, and do not bundle chainable fixes. Jones (relaying Anthropic): reduce what is reachable before a bug is found, as with the OpenSSH 10.6 change that removed code reachable before login. Both shrink the window rather than racing the finder.

Open Questions

  • No source quantifies the fix side the way the find side is quantified (bugs/day found vs. patches/day shipped at Microsoft would settle the ProPublica framing empirically).
  • Whether the Alliance produces working defensive tooling or remains a signatory list is untested — it is two days old at ingest.
  • Epoch’s discovery-vs-exploitation split suggests the asymmetry may be narrower than the headlines: if discovery on fixed budgets improved only modestly, the patching gap at Microsoft may be driven by prioritization and process, not raw find-rate.