The quantitative case against AI code quality continues to harden today, with a massive new study of 623 million commits linking AI assistants directly to an 81% spike in code duplication. Also on the desk: a textbook governance attack that legitimately drained $20 million from a DAO treasury without a single smart contract exploit, and a new vulnerability that tricks GitHub's own AI agents into leaking private repositories.
Adding to the Faros, New Relic, and SlopCodeBench datasets we've been tracking on AI code degradation, a new analysis of 623 million code changes by GitClear and GitKraken links AI tools directly to plummeting maintainability. The study found AI-assisted commits—which now account for a quarter of all commits—correlate with an 81% increase in code duplication, a 70% drop in reuse, and a 47% rise in error-masking patterns.
Why it matters
This is the largest quantitative study yet to confirm the 'structural erosion' we saw in the recent SlopCodeBench benchmark, cementing the reality that AI development velocity is currently being traded for massive long-term technical debt.
BonkDAO, the governance body for the BONK memecoin, lost an estimated $20 million from its treasury on Monday after a malicious actor acquired enough tokens to pass a governance proposal. The attacker spent approximately $4 million on BONK tokens to gain the necessary voting power to approve the proposal on the DAO's own platform, draining the funds without exploiting any smart contract vulnerability.
Why it matters
This incident is a textbook example of a governance attack, where the system worked as designed but was exploited economically, highlighting a critical risk for any DAO portal where voting power is tradable.
As teams begin deploying guardrails like Agentic OS to constrain AI coding assistants, new HKUST research shows that malicious 'skills' for tools like Claude Code and Cursor can bypass static security scanners in over 90% of cases using self-unpacking cloaking techniques. To counter this, researchers are proposing 'SKILLDETONATE,' a behavioral analysis tool that monitors the runtime actions of skills within a sandbox.
Why it matters
We've already seen 'Agentjacking' attacks trick agents into executing malicious commands, but this demonstrates that explicitly installed agent extensions can trivially evade static analysis, making runtime behavioral monitoring a hard requirement for agent environments.
A plain `CREATE INDEX` statement on a large production table takes a `SHARE` lock that can block all writes for minutes, effectively causing a write outage. The recommended approach is to use `CREATE INDEX CONCURRENTLY`, which builds the index without locking out writes, though it is slower and cannot be run inside a transaction.
Why it matters
This is a critical operational pattern for any Django developer running migrations on a live Postgres database, as failing to use the `CONCURRENTLY` option is a common cause of self-inflicted production incidents.
The same indirect prompt injection mechanism we tracked with recent 'Agentjacking' attacks—where AI assistants blindly executed malicious commands embedded in fake Sentry error logs—has now been weaponized against GitHub's own Agentic Workflows. Dubbed 'GitLost,' the new vulnerability uses a specially crafted GitHub Issue to trick the AI into executing commands that exfiltrate the contents of private repositories.
Why it matters
This reinforces that an agent's context window is a primary attack surface. Feeding untrusted external data (like issues, logs, or web pages) into privileged AI agents continues to yield critical exfiltration paths that standard repository controls cannot catch.
PayPal is phasing out its older API systems and has set a hard deadline of January 2027, after which legacy integrations will fail to process new payments. Partners like MemberPress are urging users to migrate to the new PayPal Complete Payments gateway, which supports modern payment methods and automatic refund syncing, to avoid service disruption for new sign-ups.
Why it matters
This is a mandatory, high-impact migration for any system using PayPal, as failure to update will result in silent payment failures for new customers after the deadline.
AI-Assisted Coding Is Quantifiably Degrading Codebase Health New research from GitClear and GitKraken analyzing 623 million code changes finds that AI-assisted commits correlate with an 81% rise in code duplication, a 70% drop in code reuse, and a 47% increase in error masking. This data provides the first large-scale evidence that AI coding tools are creating a significant, measurable technical debt crisis by prioritizing plausible output over maintainable, high-quality code.
Governance Attacks Emerge as a Primary Threat to DAOs A malicious proposal drained BonkDAO's treasury of $20 million by legitimately acquiring enough voting tokens to pass, bypassing smart contract security entirely. This incident, along with an urgent warning from Helius about a similar attack in progress, shows that the governance process itself is a critical and exploitable attack surface, shifting focus from code vulnerabilities to economic and procedural security.
AI Agent Security Moves to Runtime Behavioral Analysis With new research showing malicious AI 'skills' can easily bypass static scanners using cloaking techniques, the security focus is shifting to runtime monitoring. At the same time, the 'GitLost' vulnerability demonstrates how indirect prompt injection can trick agents into leaking private repo data. Both findings indicate that securing AI workflows requires observing agent behavior in a sandbox, as static analysis of code or skills is no longer sufficient.
What to Expect
Oct 31, 2026—Python 3.10 reaches its official end-of-life, after which it will receive no security patches.
Jan 2027—Hard deadline for migrating from legacy PayPal API integrations to the new PayPal Complete Payments gateway.
— The Staff Safety Desk
🎙 Listen as a podcast
Subscribe in your favorite podcast app to get each new briefing delivered automatically as audio.
Apple Podcasts
Library tab → ••• menu → Follow a Show by URL → paste