A researcher finds a WordPress RCE for $500,000 with GPT-5.6 for $25

In plain terms. A security researcher (slcyber) claims to have found a remote code execution (RCE) on a WordPress component by prompting GPT-5.6 for approximately $25 in tokens. This type of vulnerability is traded around $500,000 among zero-day brokers. The market premium to discovery cost ratio is now disproportionate - and this is a concrete case that fuels the "controlled access to frontier models" thesis.

Context

The market for critical exploits on large-scale ecosystems (WordPress ~40% of the web) has been structured for years: Zerodium, Crowdfense, and a handful of private brokers publicly quote six-figure bounties for RCEs. Automated fuzzing on PHP code is not new. What is new: the reasoning of an LLM on complex call chains and cross-cutting invariants, precisely where deterministic tools stumbled. The slcyber demonstration comes the same week as the implementation of mandatory hardware passkeys by OpenAI Trusted Access for Cyber (September 1, 2026) - the timing is no longer coincidental.

The Data

• Market premium claimed in the article: ~$500,000 (historical base Zerodium/brokers on critical WordPress RCE).
• OpenAI cost of the claimed session: ~$25.
• Model: GPT-5.6.
• Technical publication: slcyber.io, 2026-07-20.

Analysis

Three signals combine. First, the variable cost of LLM-assisted auditing falls below that of a monthly subscription - not below the threshold of serious human effort. Second, the WordPress surface remains structurally porous: thousands of plugins, random patch cycles, cross-cutting dependencies. Finally, demand from brokers does not adjust to supply in the short term: as long as the exploitation window remains open, the premiums hold. The direct consequence for frontier access control: the "no offensive security" usage policy becomes ineffective - a researcher with a standard key does what they did yesterday with a pro cybersec budget.

Scenarios

• Central (60%): The discovery remains isolated over 30 days but triggers a wave of comparable demonstrations; brokers quietly lower their premiums on "LLM-tractable" vulnerabilities.
• Positive (25%): LLM vendors harden their usage policies on "offensive code" patterns and require specific clearances; assisted discovery shifts mostly to legitimate bug bounties.
• Negative (15%): Automation lowers the barrier for private/state actors and CMS patching does not keep up - peak of massive compromises.

Implications

For a CISO, the question is no longer "who can find me an RCE" but "at what cost". For a plugin publisher, internal LLM auditing becomes a hygiene factor, not a plus. For AI platforms, the "usage policy" logic gives way to the "hard access control" logic - hardware key, jurisdiction, verified entity (fil frontier-access-control).

To Watch

• Automattic's response and potential publication of a CVE.
• Brokers' price adjustments on "LLM-tractable" segments.
• Possible extension of an OpenAI-type Trusted Access device to Anthropic / Google.