GPT-5.3 Codex Collapses Mid-Task—Builders Scramble for Fixes

The story of the day

OpenAI’s GPT-5.3 Codex—long favored for its concise, deterministic outputs—now fails mid-task without recovery, according to developer reports on Reddit. Users describe the model halting after partial responses, requiring manual re-prompting to resume. The issue persists even in OpenCode wrappers, suggesting a backend degradation rather than API-layer instability.

For builders relying on Codex for structured workflows (API stubs, config generation, or CI/CD scripts), this isn’t just an annoyance—it’s a blocker. The workaround? Downgrading to GPT-5.5, which devs describe as "verboser but stable." No official acknowledgment from OpenAI yet, but the pattern mirrors last month’s ChatGPT outage: silent failures cascading through dependent tools.

Why it matters: Codex was the rare LLM that reduced cognitive load for engineers. Its erosion pushes teams back toward either brittle prompts or human review loops—exactly what AI was supposed to eliminate.

New tools & launches

Loopsy — Local terminal-to-terminal messaging for AI agents, with Cloudflare Worker bridges for remote access. Open-source, E2E encryption in progress. [HN]

AgentShield — DeepMind’s framework for detecting adversarial prompts in agentic workflows. Focuses on jailbreak persistence across tool chains. Proprietary, research preview. [Google DeepMind]

MacOS Codex Bridge — Unofficial CLI tool to grant GPT-5.3 sandboxed filesystem access (read-only). Requires manual entitlement approvals. [GitHub]

Models & research

Agent Hijacking Taxonomy (DeepMind)

Researchers formalized 12 attack vectors for compromising AI agents, from prompt injection to "toolchain poisoning" (e.g., corrupting a agent’s external API calls). Key finding: 80% of tested agents failed to detect multi-step exploits where initial prompts appeared benign. The paper proposes runtime monitoring via "behavioral fingerprints"—but admits this adds 18-22% latency. [arXiv:2604.04321]

GPT-5.3 Codex Degradation (Community Reports)

No official benchmarks, but devs report:

• Success rate: ~68% completion on first try (down from ~92% in March)
• Recovery time: 3-5 re-prompts required for full output
• Pattern: Failures cluster around codeblocks >100 lines or nested JSON structures

OpenAI’s silence suggests either a silent A/B test or unacknowledged infrastructure strain.

Industry pulse

Chinese Courts Block AI-Driven Layoffs

Caixin Global reports that Chinese labor courts now require employers to prove AI replacements improve productivity—not just cut costs—to justify dismissals. The ruling, tested in Guangdong and Shanghai, cites "algorithmic discrimination risks" and mandates 6-month retraining periods for displaced workers. Impact: Multinationals like Foxconn and TSMC must now document AI’s net efficiency gains, not just deployment. [Caixin]

Meta’s MacOS Keylogger Backlash

After reports that Meta logs employee keystrokes for LLM training (Gradient, 4/22), Apple’s WWDC leaks suggest macOS 15 will flag unauthorized input monitoring as "privacy violations"—triggering App Store rejections. Irony: Meta’s internal tools may soon be blocked by the OS they target. [9to5Mac]

Stat of the day

87% of AI agent breaches start with a legitimate toolchain action (e.g., a Slack bot fetching a file), per DeepMind’s adversarial study. The attack succeeds when the agent’s next step isn’t constrained by its original task scope.

Get the full analysis — subscribe to Gradient Pro.